Follow upstream PocketTTS default temperature (0.3 for English models)

[adambenhassen]

Motivation

Upstream pocket-tts changed the English models' default temperature from 0.7 to 0.3 in kyutai-labs/pocket-tts@d108410 ("Human evaluations consistently prefer the English model at temperature 0.3 over the current default 0.7"). audio.cpp hardcodes 0.7 for all languages and ignores the new default_temperature yaml key, so English output at default settings no longer matches reference quality.

Changes

• PocketTTSModelConfig gains default_temperature (fallback 0.7, matching upstream's Config.default_temperature schema default for custom yamls)
• Config yaml path reads upstream's default_temperature key when present
• Builtin defaults use 0.3 for english and english_2026-04, mirroring the two configs upstream changed; other languages stay at 0.7
• Session request builders seed the request from the model default before option parsing, so an explicit temperature= option still overrides
• Generation options now also accept pocket_tts.-prefixed spellings (following the existing omnivoice.mem_saver/mem_saver alias pattern): the warm bench passes namespaced keys the session never read, so its generation options (including parity seed and noise file) were silently ignored - masked until now because the harness default equaled the hardcoded 0.7

Verification

scripts/build_metal.sh --target audiocpp_cli   # compiles clean

With language=english and no temperature option, generation runs at 0.3; temperature=0.7 (either spelling) reproduces the previous output. Non-English languages unchanged.

[0xShug0]

@adambenhassen Merged! Thank you!