Document the inplace fusion argument
`inplace` had no documentation anywhere, including the part a user most needs to
know: it drops to_q/to_k/to_v from the state_dict in favour of a single to_qkv
entry that `from_pretrained` cannot load, so the model has to be unfused before
saving. Both `fuse_qkv_projections` and `fuse_projections` now describe it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
