The AR decode loop captured the per-layer graphs without decode buffers, so
Llama2_ fell back to x = x.clone() and every replay read the address of the
first step after it had been freed. The output was garbage audio codes with no
error: the decoded track is NaN, saved as a constant -32768 by the FLAC encoder.

YuE2 and the generic generate() already pin the hidden state and rotary tensors
through decode_buffers; do the same here, gated the way generate() gates it.

Fixes #16002 (the silent-noise report) and #16222.