crispr subcommand: use multi-pattern search (v2/search_encoded_patterns) for multi-guide inputs

[nh13]

The crispr subcommand currently loops per-pattern via search_with_fn (bin/crispr.rs L227), but could use search_encoded_patterns (the v2 path used in run_batch_v2 in grep.rs) to batch all guides into a single SIMD pass per text chunk, which would dramatically improve throughput for multi-guide CRISPR off-target screening workloads.

[RagnarGrootKoerkamp]

Yeah .... so this is tricky when PAM errors are not allowed.

For forward matches, we find all end positions with sufficiently small cost, and then only trace those where the PAM has an exact match.

For rc matches, we currently search the pattern in the rc text, which is still fine.

In v2 though, we search the rc pattern in the fwd text, which means that the PAM is at the start of the match instead of at the end, and we can't currently force the trace to have matches at the start (since it greedily prefers matches at the end).

I'll try to code up a workaround where we use the v2 kernel twice in forward mode, but once with the text manually reverse-complemented.

How many guides do you have? Only if <30 or so might this workaround be slower, but otherwise it's probably faster.

[nh13]

I have say 200,000 routinely 😅