fix: recover VCF INFO fields around malformed tokens

[mwiewior]

Summary

• Replace per-field info.get() calls with a single info.iter() pass that collects all successfully parsed INFO fields into a HashMap
• Fields on both sides of malformed tokens (e.g. ;;) are now recovered instead of being silently nullified
• Removes noisy eprintln! stderr messages that fired once per record per failing field
• Also improves performance by scanning the INFO string once instead of N times per record

Context

Real-world VCFs such as Ensembl variation files contain records with double semicolons (;;) in the INFO column:

dbSNP_156;TSA=SNV;E_Freq;E_gnomAD;;MA=C;MAF=0.123;MAC=19;AA=G

The previous code called info.get(&header, &field_name) for each defined INFO field. Under the hood, noodles' get() scans the raw INFO string from the beginning each time and aborts at the first tokenization error. This caused:

1. All fields after ;; (e.g. MA, MAF, MAC, AA) to return null
2. All fields not present in the record (e.g. COSMIC_101, ClinVar_*, CLIN_*) to also error, since the scan would hit ;; before reaching the end
3. Noisy stderr output: Error parsing INFO field: "..." for every affected (record, field) pair

The fix uses info.iter() which advances past malformed tokens and continues parsing. All parseable fields are collected in a single pass, then looked up by name.

Test plan

• Added test_push_vcf_record_with_double_semicolons_in_info — verifies fields before and after ;; are both recovered (not null)
• All 106 existing tests pass
• Clippy clean (-D warnings)
• Manual test with Ensembl homo_sapiens-chr1.vcf.gz

🤖 Generated with Claude Code

[nvictus]

It's unfortunate that this requires parsing all the fields, but it doesn't look like noodles provide an alternative to tokenize without parsing.

[nvictus]

Overall, this looks good to me. Thank you!

One issue is that we now swallow tokenization and parse errors as null values. I agree that the eprintlns were noisy, but just a mental note that it would be good to provide a way to emit something in the future, or perhaps use a sentinel value to distinguish "no value" from "failed to parse".