Expand description
TiRex inference on the Burn backend (Flex CPU) for the two tensor
stages, with all host kernels reused verbatim from super.
TiRex is already ~90% engine-agnostic host code (sLSTM recurrence,
patching, scaling, SIMD kernels). Only residual_block_forward and
ffn_forward touch tensors, so this port re-implements just those two
on Burn and keeps every host helper in this module.
Weight loading reads GGUF straight via zsfm_burn::GgufWeights.