Word-level timestamps from the RNNT emission trace

[pantafive]

Implements the word-level timestamps roadmap item, in two commits: a packaging fix and the feature itself (with unit tests and README docs).

Word timestamps

RNNTGreedyDecoder already records the encoder frame index of every emitted token in RNNTGreedyStepTrace — this PR groups that existing trace into words, so timestamps cost no extra inference:

• GigaAMProTranscription and GigaAMResult gain words: [GigaAMProWord] (text, start, end in seconds) — available both from the low-level model API and from GigaAMRecognizer.transcribe.
• Tokens group at sentencepiece ▁ boundaries; punctuation pieces emitted by the e2e model attach to the preceding word; <unk> becomes a standalone ⁇ word, matching how decode renders it in the text.
• Word start = first token's frame, end = last token's frame + 1. Frame duration = mel hop (10 ms) × 4 conformer subsampling = 40 ms, exposed as GigaAMProModel.encoderFrameDuration (derived from the mel config and manifest sample rate, not hardcoded).
• SentencePieceTokenizer.piece(forTokenId:) added for piece lookup without a full decode.
• The grouping is a pure function with a new Tests/GigaAMKitTests target covering boundaries, punctuation, <unk>, blank/control skipping, mid-word chunk starts, and the frame math (swift test: 7/7). README gets a «Пословные таймстемпы» section and the roadmap checkbox.

This mirrors the technique the official gigaam python package uses for transcribe(..., word_timestamps=True) (salute-developers/GigaAM#65).

Validated in a real macOS app on Russian call recordings: timestamps line up with Silero-VAD segment boundaries, and the text output is unchanged (words are built from the same trace the text decode consumes).

Packaging fix (first commit)

Package.swift depended on mlx-swift branch: "main", which SPM refuses to resolve alongside dependents that pin mlx-swift by version — making GigaAMKit unusable as a library dependency in most real projects. Changed to .upToNextMinor(from: "0.31.3"); the API surface used here (MLX/MLXNN ops) is stable in 0.31.x. Package.resolved is regenerated to match, pinned at 0.31.3 — the newest mlx-swift line buildable below Swift 6.3 (0.31.3/0.31.4 manifests are swift-tools 5.12; 0.31.5+ require 6.3, and SPM tools-version gating lets newer toolchains resolve up to 0.31.6 while older ones stay on 0.31.x-early). Happy to adjust the range or split this into a separate PR if you prefer.

API compatibility notes

• GigaAMProTranscription's memberwise init is internal, so its added field is not source-breaking for consumers.
• GigaAMResult.init gains words with a default value — existing call sites compile unchanged. One caveat: the synthesized Codable now expects a words key, so JSON of the old shape would need a migration step if anyone persists results; say the word and I'll add a decodeIfPresent initializer.

[pantafive]

Two notes surfaced by an adversarial review while preparing this PR — both pre-existing in the repo, not introduced here, leaving them as a heads-up:

1. gigademo/.../tokenizer.model is committed as real bytes into an LFS-attributed path. Any clone with git-lfs installed shows a permanent phantom modified diff for it (the clean filter wants to turn it into a pointer), and a checkout without git-lfs would break the checksums.sha256 expectation if the blob ever becomes a pointer. Either dropping the LFS attribute for this small file or re-committing it through LFS would resolve the ambiguity.

2. README declares Swift 5.10+, but every resolvable mlx-swift version requires newer tools (0.31.3/0.31.4 → swift-tools 5.12; 0.31.5+ → 6.3). The effective minimum toolchain is 5.12 even before this PR (the previous branch: "main" pin was 6.3-only). This PR's lockfile pins 0.31.3 to keep the floor as low as mlx-swift allows.