Optimize frame reading throughput

[wtn]

Reduce per-frame overhead on the read path.

On a local benchmark (25K text frames, ~83B avg payload, Unix socket pair, IO::Stream::Buffered) throughput improved ~26%.

List of Changes

• Read both header bytes in a single stream.read(2) instead of two separate reads
• Combine extended length + mask into one read for masked frames (stream.read(6) or stream.read(12))
• Use getbyte/unpack1 instead of unpack("C").first
• Short-circuit unpack_frames for single-frame messages (to avoid map/join)

Types of Changes

• Performance improvement.

Contribution

• I added tests for my changes.
• I tested my changes locally.
• I agree to the Developer's Certificate of Origin 1.1.

[samuel-williams-shopify]

We independently verified the changes in this PR by re-applying each commit
one at a time from main, running a reproducible benchmark after every step,
and confirming the improvement before proceeding to the next.

Environment

• Ruby 4.0.2 (arm64-darwin)
• Benchmark: in-process StringIO-backed framer, median of 7 repetitions × 40,000 frames per scenario (200 frames for large payloads)
• Primary metric: median µs/frame, averaged across four common scenarios

Scenarios

Scenario	Payload	Length encoding	Masked
small unmasked	13 B	1 byte	no
small masked	13 B	1 byte	yes
medium unmasked	200 B	2 bytes (126)	no
medium masked	200 B	2 bytes (126)	yes

Large frames (70 kB, length=127 encoding) were also exercised but excluded from
the primary metric as they are dominated by payload I/O rather than parsing overhead.

Results

Step	Change	µs/frame	vs main
Baseline (main)	—	0.644	—
1	parse_header: getbyte(0) instead of unpack("C").first; integer comparisons instead of Range#include?	0.561	−12.8%
2	Framer#read_frame: read both header bytes in a single stream.read(2) call; pass second byte directly to Frame.read, eliminating a redundant read	0.449	−30.3%
3	Frame.read: unpack1 instead of .unpack(...).first; combined length+mask read for masked extended frames (one 6-byte or 12-byte read instead of two separate reads)	0.433	−32.7%
4	Connection#unpack_frames: fast-path for the common single-frame case, avoiding map + join	0.431	−33.1%

Per-scenario breakdown (baseline vs final)

Scenario	main	This PR	Δ
small unmasked	0.572 µs	0.368 µs	−35.6%
small masked	0.635 µs	0.429 µs	−32.4%
medium unmasked	0.654 µs	0.436 µs	−33.4%
medium masked	0.714 µs	0.482 µs	−32.5%

Notes

• Step 2 is the dominant win (~two-thirds of the total improvement). Eliminating
  one stream.read call per frame — by reading the 2-byte header atomically and
  passing the second byte directly into Frame.read — is the single biggest change.
• All improvements were consistent across payload sizes and mask/no-mask variants,
  indicating the gains are in parsing overhead rather than payload handling.
• The test suite was extended with 5 new cases (masked medium/large messages,
  pre-read second_byte, reserved opcode rejection) and all 92 tests pass.