feat: ft call-peaks subcommand and FIRE fiber-level filter refactor

[mrvollger]

Summary

This branch lands native peak calling in Rust, ergonomic improvements to ft pileup, a Train-FIRE Snakemake pipeline, and a handful of correctness fixes uncovered along the way.

New subcommands

• ft call-peaks — full peak-calling pipeline in Rust. Builds an FDR table per chromosome, finds local maxima, identifies natural peak start/end from underlying FIRE elements, merges adjacent peaks, and emits a BED with the expected columns. Includes lookup_fdr, reciprocal-overlap merging, and FDR builder logic with unit tests.
• ft mock-fire — synthesizes a mock BAM with FIRE elements from a BED file. Each interval becomes a FIRE element; the 4th column groups intervals into the same mock fiber. Useful for end-to-end peak-calling tests.
• ft benchmark (hidden) — micro-benchmarks the fiberseq iterator for profiling.

ft pileup improvements

• Accept multiple regions on the command line, or a BED of regions (multi-fetch instead of collect).
• Major internal refactor for code reuse with the new peak caller.
• Default frac-fibers filter; cleaner logging.

FIRE fiber-level filters → global

• --skip-no-m6a, --min-msp, --min-ave-msp-size, and the convenience --fire-filter are now global flags on FiberFilters, applied inside FiberseqRecords::next(). Every subcommand inherits them; per-subcommand duplicates are removed.
• --fire-filter defaults all three to the FIRE pipeline's values (skip-no-m6a + 10 + 10); explicit flags override (require_equals so --skip-no-m6a=false doesn't swallow the positional BAM arg).

Correctness fixes

• fa:Z extras pairing on overlapping peaks — when fibertig peaks share a forward-strand start (e.g. fs:B:I,0,0,12,159), the aligned-coordinate sort no longer reduces to "reverse the input order," so the post-hoc ann_vals.reverse() from PR fix: pair fa:Z extras with correct annotations on reverse-strand reads #103 mispaired extras. Carry extras through the flip + sort inside FiberAnnotations::new_with_extras so pairing survives any permutation. Validated on Anna's surjected CHM13 CRAM and locked down with a regression test using her exact peak shape.
• --skip-no-m6a=false override was silently ignored when combined with any other fire-filter, contradicting the CLI help. The empty-m6a reject in passes_fire_filter was hardcoded; now gated on resolved_skip_no_m6a(). The n_msps == 0 guard is kept (still required to avoid a divide-by-zero in the average-MSP-size calc).

Train-FIRE Snakemake pipeline

• New Train-FIRE/ directory with a complete Snakemake workflow for training FIRE models (rules for features, regions, apply, trackhub, train), pixi env, default profile, test data, and example mixed-positives resources (DHS/CTCF/hotspots/peaks).

Build / dev-loop

• Thin-LTO and tuned test profile (opt-level = 0 for the dev compile loop, opt-level = 2 restored where ML inference tests need speed).
• Optimize a handful of test deps so tests stay fast.
• Clippy clean (cargo clippy --tests -- -D warnings).

Tests added

• tests/call_peaks_test.rs — peak-calling integration coverage (FDR lookup, reciprocal overlap, FDR builder).
• tests/fibertig_test.rs — fa-pairing tests for forward, reverse-strand multi-peak, and reverse-strand shared-start cases.
• src/utils/input_bam.rs — 7 unit tests for the new fire-filter resolution and passes_fire_filter semantics (each threshold, --fire-filter combo defaults, explicit overrides).

Test plan

• cargo test — 70+ tests across unit + integration suites
• cargo clippy --tests -- -D warnings — clean
• ft pileup smoke check (baseline 13192, --fire-filter → 12967, --min-msp=1000 → 1)
• ft fire smoke check
• ft call-peaks smoke check
• ft pg-inject --extract on Anna's surjected CHM13 CRAM