Simplify stream's pcmf32 handling

[tamo]

Use one deque instead of two vectors (old and new).
Old and new are length variables now.

Basically: Get step - new samples every time.
Then substitute new = (around) step;
The new audio data is simply appended to the deque.
(Limit the deque size to 30 seconds.)
Pass old + new samples to whisper inference.

If the data has been consumed, let old = 0; new = 0;
If some of the data should be kept for the next iter, old = keep;
If you want to get only N samples next time, new = step - N;

In VAD mode: stream --interim --step -3000 will
Get 3000ms of audio.
Run vad_simple(step_ms).
If nothing is detected, get 100ms more audio and retry.
If nothing is detected and 3000ms has been passed,
go into the interim mode,
where n_segments - 1 segments will be confirmed.
(old -= confirmed_t1)
If n_segments == 1, only show the first half of the result.

Misc:
Increase the default max_tokens because 32 is too small for 10 seconds.
(Some Japanese speech was garbled.)
Write wav as soon as the data is available.

no_timestamps is the default even for VAD
because it is more useful to show to the hard-of-hearing

[tamo]

Closing, because a number of fixes are in #2694 now.