perf: replace difflib fuzzy aligner with O(n*m^2) LCS DP

[aksg87]

Fixes #386, fixes #197. Related: #188.

Summary

Replace the legacy fuzzy aligner with an O(n·m²) time, O(m²) memory
LCS DP gated by coverage and density rules, where n is source length
and m is extraction length (typically 3-8 tokens). The old path
enumerated candidate windows and scored each with difflib, which
scales badly on long documents.

Speedup: 1,000-token sources drop from ~33s to under 5ms; 5,000-token
sources from ~69min (est.) to under 20ms.

Changes

• LCS DP with rolling rows: O(m²) memory regardless of source length,
  no allocation guard needed
• Tightest span tracked per achievable match count, so a denser
  sub-match can be considered when the max-match span fails density
• Coverage and density acceptance gates to reject weak partial matches
• Pre-normalized source tokens computed once per chunk
• New params are keyword-only; positional callers are unaffected
• Fuzzy-param validation and deprecation warning gated behind
  enable_fuzzy_alignment so exact-match-only callers stay quiet
• Algorithm choice is consistent between prompt validation and runtime
• Legacy algorithm kept as a deprecated opt-out with DeprecationWarning

Test plan

• 443 tests pass (39 new)
• Covers LCS correctness, acceptance gates, sparse-vs-dense submatch
  regression, positional API compat, deprecation warning, param
  validation, fuzzy-disabled quiet path, and end-to-end extract()
  integration