Support the legacy single-function SpeechDecoder and MultiCodeDecoder assets

[EduardoPach]

#494 and #513 moved the SpeechDecoder and the MultiCodeDecoder onto multifunction
assets, which dropped support for the legacy single-function W8A16 ones. This
restores both by detecting the layout at load time instead of assuming it.

SpeechDecoder

• Probe MLModelAsset.functionNames; empty → load without a functionName.
  Requesting throughputOptimized against it errors instead of silently
  producing latency-mode audio.
• Read the kv_cache_update_mask rank from the model (2 = legacy, 3 =
  multifunction). Not derivable from codesPerStep — both ranks exist at
  codesPerStep == 1, so SpeechDecoderCache takes it as a parameter.
• ModelUtilities dimension reads are now bounds-safe; probing an
  unexpected-rank input used to trap.

MultiCodeDecoder

Same probe: the legacy asset is schema-identical to the multifunction stepped
graph, so it loads without a functionName and drives the same decode path.
.fused has no counterpart in a single-function asset, so asking for it reports
a configuration error naming the multifunction variant rather than quietly
running stepped. .stepped stays the default; both variants are supported.

Downloader

setupModels built a fresh TTSKitConfig for the download that dropped
versionDir and the six component variants. downloadPatterns is derived from
those, so any pinned variant was ignored — the downloader fetched the preset
defaults and the load then failed on the variant actually requested.

This is pre-existing (it reproduces on main), but it makes the legacy support
above unreachable through TTSKit(config:), so it is fixed here. It also means a
user who already has W8A16 and pins it currently pulls the ~1.1 GB
multifunction assets they don't need before failing to load; after the fix that
run downloads nothing.

Validation

Cache deleted and re-downloaded from scratch: all 40 files match their original
SHA-256. Same text/seed at --temperature 0, vs main:

comparison	result
multifunction stepped / fused / throughputOptimized vs main	bit-exact
legacy W8A16 vs multifunction stepped	corr 0.99999, max abs 7.4e-3
pinned W8A16 against a cache that already has it	0 bytes downloaded

Both guards fire: legacy + fused and legacy + throughputOptimized each report
the single-function configuration error. 4 tests added for rank-2 geometry. The
macOS 14 path stays deleted — both layouts need the macOS 15 / iOS 18 MLTensor
floor.

One commit is unrelated to the decoders: it gates xcbeautify --renderer github-actions to a single matrix job, since eight jobs were annotating the same
warning eight times.

[EduardoPach]

This PR will allow people to do the following in case they have the previous
assets already downloaded and don't want to download the new
W8A16-multifunction ones:

import TTSKit

Task {
    let config = TTSKitConfig(
        multiCodeDecoderVariant: "W8A16",
        speechDecoderVariant: "W8A16"
    )
    let tts = try await TTSKit(config)
    let result = try await tts.generate(text: "Hello from TTSKit!")
    print("Generated \(result.audioDuration)s of audio at \(result.sampleRate)Hz")
}