Fix pileup indel support: correct D/I CIGAR inversion and bounds guard

[pkerpedjiev]

Summary

• D/I CIGAR inversion (clodius/tiles/pileup.py): nw_trace_scan_profile_16 builds the alignment profile from the reference, so the returned CIGAR describes the alignment from the profile's perspective — its D and I ops are swapped relative to SAM convention. tile_functions_parasail now normalises the CIGAR before passing it to cigar_to_subs, so insertions and deletions are reported with the correct type.
• Bounds guard in cigar_to_subs: an X (mismatch) op that extends past the end of the reference now clips gracefully instead of raising IndexError. This was benign when all query sequences had the same length as the reference (SNPs only) but was triggered as soon as queries had variable lengths.
• sequence_pileup.py benchmark: updated mutate_sequence to apply Poisson-sampled insertions (~3) and deletions (~3) per sequence in addition to SNPs, and reports per-event-type counts from the alignment output.

Test plan

• TestCigarToSubs::test_x_op_clipped_at_ref_boundary — verifies the bounds guard clips X ops that would index past the reference end
• TestGetPileupAlignmentDataParasail::test_insertion_detected — a query longer than the reference produces an I event
• TestGetPileupAlignmentDataParasail::test_deletion_detected — a query shorter than the reference produces a D event
• TestGetPileupAlignmentDataParasail::test_mixed_indels_and_snps — a sequence with insertion + deletion + SNP produces all three event types (I, D, X)
• TestGetPileupAlignmentDataParasail::test_variable_length_sequences_align_without_error — a batch of sequences with different lengths all align without error and produce the expected event types
• All pre-existing pileup tests continue to pass