refactor!: replace the pipeline with a design of our own

[Halvanhelv]

Why this exists

Six files in this gem derived from
google_translate_diff:
request.rb, linearizer.rb, spacing.rb, chunker.rb, tokenizer.rb and
cache.rb. That project carries no licence at all — no LICENSE file,
nothing in its gemspec, and none reported by GitHub. All rights reserved.
This gem ships under MIT, which nobody here is in a position to grant for
code that is not theirs.

This branch deletes all six, and their tests, and replaces them with a
pipeline of our own.

How the replacement was written

From the tests and a written spec, without reading the files being
replaced. The clean-room constraint bound every commit on this branch that
wrote the new pipeline; only this final commit — the switch-over, the
deletion and the docs — was written with the old files open, and only to
check what referenced them. Nothing was copied out.

The first commit on the branch (0af12af) pinned the old pipeline's
behaviour before any of it changed: 32 corpus inputs, each a case some
earlier bug or review turned up, translated through the :null provider and
recorded to a baseline file. test/translation_diff/pipeline_corpus_test.rb
judges the new pipeline against that baseline, input by input.

A reviewer compared the two designs and found the replacement independent
work.

What replaces what

Gone	Replaced by
Linearizer	Document (walks a caller's structure without flattening it) and Leaves
Tokenizer	Passage, Fragment and Markup (markup and prose, without rebuilding either)
Spacing	Segment (a sentence keeps the whitespace it was found in)
Chunker	Batch (packs sentences into provider-sized requests)
Cache	SentenceCache
Request	Translator

TranslationDiff.translate and Context#translate keep their signatures.
docs/how-it-works.md describes the new pipeline.

What changes for a caller

Cache keys are otherwise unmoved. SentenceCache's key format is pinned
by test against values recorded from the pipeline being replaced
(test_keys_match_the_ones_the_previous_pipeline_produced and friends in
test/translation_diff/sentence_cache_test.rb), including the options
digest's canonicalisation, which is reproduced collision-for-collision rather
than fixed — changing it would move every key.

Five corpus inputs change, and they were declared in
PipelineCorpus::EXPECTED_TO_CHANGE before the rewrite, on purpose, so the
list could not be assembled after seeing what broke:

• Salt & pepper. was sent to the provider as the six characters
  & — the provider translated the entity's spelling as words, and we
  paid for it. Now sent as Salt & pepper. and re-encoded on the way out.
• Hard space here. likewise.
• if a < b then stop. Fine. — ox read the < as an unclosed tag, so
  only if a was ever translated. A < that opens no tag is now escaped
  before parsing and restored after.
• 5 < 6 and 7 > 6. True. — sent as three fragments, now sent as two
  sentences.
• a <b then stop. Fine. — a recorded limit, not a fix: <b cannot be
  told apart from a tag without a lexer of our own.

Two further key changes, both in the CHANGELOG:

• A sentence padded with Unicode whitespace (a non-breaking space, say) now
  keys as the bare sentence. The pipeline uses one Unicode-aware definition of
  padding everywhere; the old key was built with ASCII strip, which leaves
  U+00A0 in place.
• < still reaches a provider undecoded, a known limit of working around
  ox rather than replacing it.

Documents containing any of these will miss the cache once and be
re-translated. That is the point: what was cached for them was translated from
the wrong text.

Also breaking, and in the CHANGELOG:

• Request::Error → Translator::Error, Cache::Error →
  SentenceCache::Error, Chunker::Error gone (the condition it named now
  raises TranslationDiff::Error from Batch). No aliases: this gem has
  never been published under the name translation_diff with those constants
  in it. All three stay TranslationDiff::Error, so a rescue of the base
  class is unaffected.
• to: missing or nil raises ArgumentError naming the keyword. It
  previously compared equal to a nil source, short-circuited as "same
  language", and handed your values back untranslated in silence.
• The cache instrumentation event fires once per call, not once per chunk —
  the cache is now consulted for every sentence in one read_multi before
  anything is batched. hits and misses sum to the same totals, so a
  counter dashboard is unaffected; an event-rate counter or a per-chunk
  histogram will notice.

The README line

**TranslationDiff** based on [GoogleTranslateDiff](...)

is removed in this commit and no other. It was true while the code was
derived. It stops being true when the last derived file is deleted, which is
this commit.

Verification

• Full suite green on seeds 1, 7, 99, 1234 and 4242 — five, because this
  changes the path every other test runs through. 502 runs, 1176 assertions,
  0 failures, 0 errors, 0 skips.
• bundle exec rubocop — 83 files, no offenses.
• Line coverage 98.6%; nothing in the new pipeline is uncovered.
• All seven providers still register.
• No reference to google_translate_diff or the six deleted constants
  remains in lib, test, docs or README.md, outside the CHANGELOG's own
  history and the rename note in docs/errors.md.

One behaviour test edited, and why

pipeline_corpus_test.rb's CHANGED table carried a document: column
recording what the old path still produced for the five changed inputs, so
that half of each case held the line until the switch. With the old path
gone, the column is dead: the assertion now compares against echoed:, which
Task 5 wrote for exactly this moment. Only entity nbsp actually differed
between the two (  renders as the character it means — the same
document to a browser, not the same bytes). No other behaviour test was
touched.