grep: keep only the requested number of before-context lines

[MsfPablo]

Fixes #66.

ContextBuffer::new rounds its slot count up to a power of two so push can mask instead of divide. push then capped the live length with .min(self.slots.len()) — the rounded count — so the buffer retained up to the next power of two rather than the requested -B N. That is why -B 1/2/4/8 looked correct while -B 3 printed four lines.

Kept the power-of-two slot allocation (the mask in push/drain_iter depends on it) and stored the requested capacity separately to bound len.

$ printf '01\n...\n10\nMM\n' | grep -e MM -B 3
08
09
10
MM

Diffed -B and -C for N=1..8 against GNU grep on the issue's input: identical for every value. Added a regression test over N=3,5,6,7 — the non-power-of-two values, since the powers of two passed even with the bug.

cargo test: 95 passed. clippy -D warnings and cargo fmt --check clean.

[codspeed]

Merging this PR will not alter performance

✅ 10 untouched benchmarks
⏩ 17 skipped benchmarks¹

---

_{Comparing MsfPablo:before-context-exact-count (408ba6a) with main (5882b3d)}

Footnotes

1. 17 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[lhecker]

Permit my rudeness but this is identical #69 and I find it funny in a sad way that both are the exact same type of wrong. Seriously, they're line-for-line identical, almost as if... it was written by the same entity.

This is precisely why I'm jaded of AI-authored PRs: There's just... no variety anymore. They're all perfectly predictable within stochastic bounds. Same inputs expected outputs. And amnesia is built in as a feature.

PR rejected.