[MEDIUM] Free COPY buffers when decoders raise

[OskarEichler]

Summary

PG::Connection#sync_get_copy_data releases the libpq buffer after decoding, but a decoder exception bypasses PQfreemem. Repeated rejected rows therefore retain every COPY buffer and can exhaust process memory.

This change wraps decoding in rb_ensure, so the libpq buffer is released on both success and exception. Successful decoding behavior is unchanged.

Reproduction

I used a custom PG::SimpleDecoder whose decode method raises, then rescued each error while reading four local COPY streams of 32 one-MiB rows. With current master, RSS rose each round:

start:    23,264 KiB
round 1:  91,280 KiB
round 2: 124,720 KiB
round 3: 158,176 KiB
round 4: 191,632 KiB

With this patch, memory stabilized after allocator warmup:

start:    22,160 KiB
round 1:  42,064 KiB
round 2:  48,336 KiB
round 3:  57,744 KiB
round 4:  57,760 KiB

Each run called GC.start before measuring.

Verification

• Built the native extension on Ruby 4.0.6 / PostgreSQL libpq 15.
• Repeated the four-round memory reproduction before and after the patch.
• Focused COPY/type-map coverage: 18 examples, 0 failures.
• Full suite: 1,111 examples, 8 failures, 9 pending. All eight failures are the existing local pg_service/libpq 15 expectation mismatch; the COPY coverage passed.
• git diff --check passes.

Limitations

The memory measurements used a local PostgreSQL 15 server and RSS, so allocator retention accounts for the patched run's initial rise. No new automated spec is included.

Breaking changes

None.