perf: speed up read_span with bulk memchr line scanning

[connorshea]

Summary

context_info (the engine behind every built-in SourceCode::read_span impl) walked the source byte-at-a-time from offset 0 for every call. This PR scans the region before the span in bulk with memchr2 instead, deriving line count, context line starts, and column arithmetically from newline positions. The original byte-loop is kept verbatim for the span itself and the trailing context, so behavior is unchanged.

It also fixes a panic found while validating the change: a zero-length span starting exactly one byte past EOF ("a".read_span(&(2, 0).into(), 0, 0)) sliced out of bounds instead of returning Err(MietteError::OutOfBounds). This was reachable from GraphicalReportHandler, which calls read_span(label, 0, 0) on the primary label, so an off-by-one zero-length label span could panic report rendering. A regression test is included.

memchr moves from transitive to direct dependency.

Performance

Measured on a 3MB / 40k-line source, span near EOF, --release:

Metric	before	after	speedup
read_span, LF endings	2.17 ms	260 µs	8.4x
read_span, CRLF endings	2.20 ms	453 µs	4.8x
full graphical render, 1-label diagnostic	3.43 ms	390 µs	8.8x

Render cost previously scaled linearly with file size (3 full-file scans per rendered diagnostic); those scans are now ~8x cheaper. Heap usage is unchanged: one 32-byte VecDeque allocation per call at default context settings, independent of file size (verified with a counting allocator).

Verification

• Differential-fuzzed the new implementation against the old one: 309,698 cases (random content with LF/CRLF/lone-CR mixes, random spans including zero-length and out-of-bounds, context settings 0–3) with 0 mismatches; the 68 cases where the old code panics now return Err(OutOfBounds).
• Full test suite passes including the 38 graphical golden-output tests (the color_format failures are pre-existing/environmental — they fail identically on main in a non-tty sandbox).

🤖 Generated with Claude Code

[connorshea]

I need to review/do more testing before marking this ready