Fix maybe_autocast crashing on meta device tensors

[Butanium]

What does this PR do?

maybe_autocast calls torch.is_autocast_enabled(device_type) which raises a RuntimeError when device_type is "meta":

RuntimeError: unknown device type for autocast in get_autocast_dispatch_key_from_device_type

This breaks any code that runs a forward pass on meta tensors — for example, nnsight's .scan() context, which traces the computational graph using meta tensors without materializing weights.

The error path is: LlamaRotaryEmbedding.forward → maybe_autocast(device_type="meta", enabled=False) → torch.is_autocast_enabled("meta") → 💥

Since autocast is meaningless on meta tensors (they don't compute anything), this PR returns nullcontext() early when device_type == "meta".

This affects all 20+ model files that use maybe_autocast (via RoPE or directly), so fixing it at the source in maybe_autocast is preferable to patching each callsite.

Reproduction

import torch
torch.is_autocast_enabled("meta")  # RuntimeError

# With nnsight / any meta-tensor forward pass:
from nnsight import LanguageModel
model = LanguageModel("meta-llama/Llama-3.1-70B")
with model.scan("test"):  # crashes in RoPE forward
    pass

Fix

def maybe_autocast(device_type, dtype=None, enabled=True, cache_enabled=None):
    if device_type == "meta":
        return nullcontext()
    ...

Note on AI usage

🤖 Generated with Claude Code
👨 reviewed by the human:
This bug came out in some mech interp class and I maintain the nnterp library that was used and encounter this issue. claude helped me to traceback the issue to transformers and the fix proposed fix our issue and is clean and minimal

[Butanium]

@scienceetonnante would be nice if this could be merged and included in next transformers version for nnterp! This came out in @gsarti mech interp class with this nnterp snippet

model = StandardizedTransformer(
    "meta-llama/Llama-3.1-70B",
    device_map="auto",
    remote=True)

I'll also push a fix for nnterp to fallback to remote execution renaming check if scan fails but that's a bit annoying

[Rocketknight1]

This LGTM but I should check with the people who touched this code recently - cc @Cyrilvallez @hmellor for final review!

[hmellor]

Seems reasonable

[Cyrilvallez]

Weird usage pattern but alright!

[Butanium]

Thanks for the quick review and merge!