Surface CTC vocabulary rescoring decisions on the public ASR results

[kevin-nous]

Problem

With vocabulary boosting configured, VocabularyRescorer already builds a
RescoringResult for every word it decides on — the original word, the
original and replacement scores, and the reason. Nothing reaches the caller.
The public results expose ctcDetectedTerms and ctcAppliedTerms, so an app
can see which vocabulary terms went in, but not which decoded word each one
displaced or the scores behind the decision.

That makes two things impossible downstream:

• Telling a recognizer error apart from a bad vocabulary replacement when a
  transcript reads wrong.
• Showing a user what boosting actually changed ("codecs" → "Codex").

The decisions are dropped at the call sites: SlidingWindowAsrManager flattens
them to [String] via compactMap { $0.replacementWord }, and everything else
keeps .text.

Change

• ASRResult gains ctcReplacements: [VocabularyRescorer.RescoringResult]?,
  aligned 1:1 with ctcAppliedTerms.
• SlidingWindowTranscriptionUpdate gains the same field, so the streaming path
  exposes what the batch path does.
• withRescoring(text:detected:applied:) gains a defaulted replacements:
  parameter rather than a parallel helper.
• VocabularyRescorer.RescoringResult gains Codable (all stored properties
  are String/Float/Bool), because ASRResult is Codable.
• TextNormalizer.normalize(result:) carries the new field through.

The Unified path needed its own accessor: UnifiedAsrManager.transcribe(_:)
returns a String and transcribeWithTimings(_:) returns text plus timings, so
neither has anywhere to put the CTC metadata — a field on ASRResult alone
would be unreachable for callers of the offline manager. So:

• UnifiedAsrManager.transcribeDetailed(_:) returns the whole ASRResult —
  text, confidence, duration, processing time, token timings, and all three CTC
  fields. Naming follows the existing synthesizeDetailed pairs in
  KokoroAneManager / PocketTtsManager.
• StreamingUnifiedAsrManager.consumeVocabularyReplacements() drains the
  decisions applied since the previous call, mirroring the existing
  consumeTokenTimings() / consumeWordTimings() accessors; its finish()
  returns a String and cannot carry them.
• rescoreIfConfigured now returns the rescorer's RescoreOutput instead of
  discarding everything but .text.

Compatibility

Additive and source-compatible. Both new fields default to nil and are the
last parameter of their memberwise init; the new withRescoring parameter is
defaulted, so existing three-argument calls compile unchanged. transcribe(_:),
transcribe(_:AVAudioPCMBuffer) and transcribeWithTimings(_:) keep their
signatures and their behaviour, including not building token timings when
boosting is not configured. ctcDetectedTerms and ctcAppliedTerms are
untouched.

Testing

Tests/FluidAudioTests/ASR/Parakeet/VocabularyRescoringSurfacingTests.swift
covers the batch path (withRescoring carrying the decisions, the defaulted
call leaving them nil, a Codable round trip, and the pure
UnifiedAsrManager.applying(_:to:) / meanConfidence(of:isEmpty:) helpers
transcribeDetailed is built from) and the streaming path
(SlidingWindowTranscriptionUpdate carrying and defaulting them, and the
StreamingUnifiedAsrManager drain). The rescoring mapping is factored into
pure statics for the same reason tokenTimings already is: it keeps the rule
testable without loading a 600M parameter model or mocking one.

swift test: 2492 tests, 57 skipped, 1 failure. The failure is
LuxTtsG2pTests.testFixtureResourcesAreProcessedAtBundleRoot, which fails the
same way on an unmodified main checkout here and is unrelated to this change.
swift format lint reports nothing new on the touched files.