Phonon 0.5.0

Changes since 0.4.2.

• Small models: the runtime loads the architectures of Kyutai's small Pocket TTS models: one-step flow heads (flow_type: drifting or pointwise), a z_dim projection, grouped-query attention (attn_dim, num_kv_heads) and a Mimi encoder inner_dim. Every new key is optional, so existing checkpoints load and sound the same. The Core ML export does not support these keys yet and says so (#210).
• Mimi: the decoder's streaming attention now covers context positions, as in the Python model the decoders were trained with, instead of one more. Output changes slightly, toward the reference (#209).
• Fixes:
  • Browser: phonon-tts accepts the abbreviations and elongations rewrite rules (#274).
  • Voice catalogs: adding a catalog keeps the checkpoint's own voices, and a local catalog's voice files must be inside its folder (#275).
  • Python: the type stub lists normalize in __all__ (#256).
  • Swift: a locally built PhononCore.xcframework targets each platform's default CPU, as release builds already did (#257).
• Rust API (breaking): ConvDownsample1d::load and StreamingMultiheadAttention::load take new arguments, and public config structs have new fields (#210). This is why the version is 0.5.0.
• Housekeeping: unused dependencies and dead code removed (#258, #259, #264), crates that are not on crates.io marked publish = false (#265), both licences declared in the Android QNN library's POM (#268), and doc fixes (#270, #271, #272).

The attached model bundles (phonon-31m-en-coreml.zip, phonon-31m-en-qnn.zip) are those of 0.4.0: nothing in this release changes what they contain.