[mxfp8 training] add TP warning

[danielvegamyhre]

Summary

• We are doing a refactor of how Dtensor + MXFP8WeightWrapperTensor subclass compose, reversing the wrapping order to MXFP8WeightWrapperTensor(Dtensor(...)) ([TP] reorder MXFP8 wrapper over DTensor ao#4010) and landing Dtensor scaled_mm sharding strategy fixes ([DTensor] fix scaled_mm sharding strategy pytorch#177234).
• In the meantime, while the wrapping order is Dtensor(MXFP8WeightWrapperTensor(..)), we need to warn that linears using TP will use default precision, not MXFP8.

Optional extra details

• Due to how Dtensor linear uses "composite implicit autograd" it decomposes linear ops into aten.t + aten.mm, going straight through __torch_dispatch__ instead of first going through __torch_function__. This prevents our subclass from intercepting the linear op to dispatch to _to_mxfp8_then_scaled_mm autograd func. We cannot intercept at the __torch_dispatch__ level because then autograd would not capture the backward for our autograd func we dispatch to.
• In contrast, grouped_mm does not have this problem, as it has an explicitly registered backward (compose explicit autograd) so we always see aten._grouped_mm in __torch_function__ and can intercept.
• thanks @pianpwk for the help with this!