[full_dtensor] Allow sharding config to wrap declared plain inputs

[Matrix-Z97]

Summary

This PR updates Module._redistribute_inputs() so plain Tensor inputs with explicit in_src_shardings are wrapped via DTensor.from_local() before the full_dtensor plain-Tensor guard is applied.

With this change:

• plain Tensor inputs declared in in_src_shardings are converted according to their sharding config;
• undeclared plain Tensor inputs still raise in full_dtensor mode;
• existing DTensor placement validation remains unchanged.

Why

The existing ShardingConfig contract documents that in_src_shardings can be used to annotate plain tensors as DTensors at module boundaries.

However, the previous full_dtensor logic rejected any plain Tensor input before checking whether that input had a declared source sharding. As a result, this documented path was unreachable in full_dtensor mode:

plain Tensor + in_src_shardings -> DTensor.from_local(...)

I am not entirely sure whether the previous behavior was intentional or a bug. The current ShardingConfig and _redistribute_inputs() documentation both suggest that plain Tensor inputs with explicit in_src_shardings should be wrapped at the module boundary, but full_dtensor mode previously raised before that path could run. Could reviewers confirm whether this relaxation matches the intended full_dtensor contract?

[pytorch-bot]

Workflows were awaiting approval. CI has now been triggered for the ciflow labels on this PR.

[fegin]

For full dtensor mode, the inputs shouldn't have plain tensors. The raise is intentional, not a miss. in_src_shardings for plain tensor use case is primarily preserved for the legacy mode.