Whisper speech-recognition models (tiny, base and small, multilingual), ONNX format

Weights: OpenAI Whisper, https://github.com/openai/whisper
         "Whisper's code and model weights are released under the MIT License"
         (LICENSE-whisper-MIT.txt, copied from the openai-whisper 20250625 source
         distribution on PyPI).
ONNX export: k2-fsa/sherpa-onnx (Apache License 2.0, LICENSE-sherpa-onnx-Apache-2.0.txt),
         GitHub release "asr-models":
           https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-tiny.tar.bz2
             SHA-256 c46116994e539aa165266d96b325252728429c12535eb9d8b6a2b10f129e66b1
           https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-base.tar.bz2
             SHA-256 911b2083efd7c0dca2ac3b358b75222660dc09fb716d64fbfc417ba6c99ff3de
           https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-small.tar.bz2
             SHA-256 486a46afbb7ba798507190ffe02fea2dd726049af212e774537efac6afb210a6

Files and modifications:
- tiny-encoder.onnx: tiny-encoder.int8.onnx, unmodified.
- base-encoder.onnx.part0..1: base-encoder.int8.onnx, unmodified, split into two byte ranges.
- tiny-decoder.onnx.part0..1, base-decoder.onnx.part0..2: {tiny,base}-decoder.int8.onnx with
  one change: the float32 token-embedding matrix (textDecoder.token_embedding.weight) is
  stored as int8 with one float32 scale per row and restored by a Cast + Mul at the start
  of the graph (maximum absolute error 0.0012). This shrinks the decoders from 90 MB / 131 MB
  to 30 MB / 51 MB. The files are then split into byte ranges of at most 20 MB so each one
  fits static-hosting per-file limits; the browser concatenates the parts before loading.
- small-encoder.onnx.part0..6: small-encoder.int8.onnx with its three float32 initializers over
  0.5 MB (the two convolution kernels and audioEncoder.positional_embedding) stored as int8 with
  one float32 scale per output row and restored by Cast + Mul (maximum absolute error 0.0039);
  112 MB -> 103 MB. Split into 7 byte ranges of at most 17 MB.
- small-decoder.onnx.part0..8: small-decoder.int8.onnx with the token-embedding change above
  (maximum absolute error 0.0012); 262 MB -> 143 MB. Split into 9 byte ranges of at most 17 MB.
  Checked with Python onnxruntime 1.25 (one thread): the same greedy transcripts as the
  unmodified int8 files on English (LibriSpeech test_wavs/1.wav) and Arabic test audio.
- tokens.txt: tiny-tokens.txt (identical to base-tokens.txt), unmodified: the Whisper
  byte-level BPE vocabulary as "base64(token bytes) id" lines.

scripts/copy-whisper.mjs copies these files to /public/whisper/, where the transcription
tools load them with ONNX Runtime Web (MIT, Microsoft), which is served from /bg-removal/
by scripts/copy-bg-removal.mjs. Nothing is downloaded from third-party hosts at runtime.