mtp: make speculation disable skip draft work

[abhay]

Summary

DS4_MTP_SPEC_DISABLE should leave greedy generation on the ordinary
non-speculative path. Before this change, the flag bypassed speculative
acceptance, but two pieces of speculative plumbing still ran:

• one-shot CLI generation still selected the session sampling path when MTP was
  configured
• ds4_session_eval() still prepared MTP draft tokens after each decode step

This patch honors the flag at those remaining decision points. DS4_MTP_PROBE
remains independent, so explicit draft-probe diagnostics still work without
enabling speculative generation.

Why This Is A Correctness Fix

The disable flag previously meant "do not accept speculative tokens" rather
than "do not run speculative work." That made the disabled mode surprising and
kept extra MTP work in a path meant to be ordinary greedy generation. After this
change, the flag has the expected behavior: configured MTP support can remain
loaded, but speculative generation work is skipped unless speculation or probe
diagnostics are explicitly enabled.

Testing

Machine/backend/model:

• Apple M5 Max, Metal backend, 128 GiB RAM
• main model: DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix.gguf
• MTP model: DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf

Commands run on this branch:

make clean
make
make test

make passed.

make test ran with Metal access. The long-context, tool-call-quality,
metal-kernels, and server tracks passed. The logprob-vectors track failed
on long_memory_archive with 7 failures.

I checked the same logprob-vector track on main:

DS4_TEST_MODEL=/Users/abhay/Dev/ds4/ds4flash.gguf ./ds4_test --logprob-vectors

It failed the same long_memory_archive vector on main with the same 7
failures, so this branch does not appear to introduce that regression. That
pre-existing vector failure is addressed separately in
#204, which runs the official logprob-vector
check through the quality path.

Targeted Disabled-MTP Smoke Check

I also compared the exact affected mode against main:

env DS4_MTP_SPEC_DISABLE=1 ./ds4 \
  -m /Users/abhay/Dev/ds4/ds4flash.gguf \
  --mtp /Users/abhay/Dev/ds4/gguf/DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf \
  --mtp-draft 2 \
  --temp 0 \
  --nothink \
  --ctx 1024 \
  -n 32 \
  -p 'Write one concise sentence about testing.'

Both main and this branch produced the same greedy output. Generation speed in
this narrow smoke check was:

• main: 35.67 t/s
• this branch: 38.98 t/s

This is not a broad throughput claim. It only shows that the disabled-MTP mode
now avoids the extra draft-path overhead this patch removes.