docs: add custom ASR shim proxy example

[xAlcahest]

Summary

OpenWhispr's "Cloud → Custom" transcription assumes the OpenAI wire format: it POSTs multipart/form-data to {BaseUrl}/audio/transcriptions and reads back {"text": "..."}. ASR APIs with a different shape (StepFun StepAudio 2.5, and several Chinese cloud providers) fail with 400/501-type errors and no obvious guidance. This adds an examples/custom-asr-shim/ directory with a small local proxy that bridges the two without touching OpenWhispr: the shim listens on /audio/transcriptions, transcodes the recording with ffmpeg, calls the vendor in its own format, and returns the OpenAI-shaped JSON. shim_template.py is a vendor-neutral skeleton where you fill in one transcribe() function; stepaudio_shim.py is a working StepAudio 2.5 implementation ported from @ErogosZhou's gist.

Closes #972

Changes

• examples/README.md: index for the examples directory.
• examples/custom-asr-shim/README.md: the guide — the wire protocol OpenWhispr speaks, the HTTP-vs-HTTPS private-host rule, prerequisites, quick start, and how to adapt it to another vendor.
• examples/custom-asr-shim/shim_template.py: vendor-neutral shim (ThreadingHTTPServer bound to 127.0.0.1, stdlib multipart parser that avoids the removed cgi module, ffmpeg transcode, 25 MB body guard, temp-file cleanup).
• examples/custom-asr-shim/stepaudio_shim.py: StepAudio 2.5 shim (raw PCM base64 request, SSE response parsing).
• examples/custom-asr-shim/test_shim.py: unittest coverage for the multipart parser, the SSE parser, and the HTTP endpoints.
• examples/custom-asr-shim/.gitignore: ignore Python bytecode.