Remove redundant audio decoding in CLI transcribe path

[hamzaq2000]

Why is this change needed?

fluidaudiocli transcribe currently decodes batch audio three times: once into an unused buffer, once during unconditional resampling, and once inside AsrManager.transcribe.

try audioFileHandle.read(into: buffer)                           // Full decode #1
let samples = try AudioConverter().resampleAudioFile(...)        // Full decode #2
let result = try await asrManager.transcribe(audioFileURL, ...)  // Full decode #3

The decoded buffer was never used, so that read is removed. The second decode only supplied the sample-count log (removed) and custom-vocabulary samples. The CLI now passes normal files directly to AsrManager; with custom vocabulary, it decodes once and shares the samples between ASR and CTC.

An existing long-form duration bug was also fixed, alongside the above main patch. ChunkProcessor passed an empty sample array when building its result to avoid materializing disk-backed audio, which produced a zero duration. The result builder now accepts a sample count instead, allowing chunked paths to pass their existing totalSamples value.

Validation

Benchmarked Parakeet V3 on Apple M3 using a 1 hr mp3:

	Upstream	PR branch
Time	44.09 s	18.50 s
Peak memory	2,664 MiB	354 MiB

Yields a ~2.4x speedup with an 87% reduction in peak memory. The resulting transcript was byte-for-byte identical.

The equivalent WAV test showed no performance regression (15.29 s vs. 15.07 s) and also produced an identical transcript. A custom-vocabulary test likewise produced byte-identical output while improving from 2.50 s to 1.75 s.