Fix pytree input validation and use solved placements for forward validation

[fmassa]

The simple API (auto_parallel) advertised pytree support (dict inputs, nested structures) but had two bugs that broke it at runtime:

1. _compute_expected_inputs iterated over traced_inputs directly, so a dict element hit the non-tensor branch and sharding was never applied to its tensor leaves. The forward function similarly passed the raw pytree structure to the compiled graph, which expects flat tensor args.
2. Forward validation used self.input_constraints (the user's request) rather than the solver's solution to determine expected input shapes. If the user set no constraints and the solver chose something other than the hardcoded default Shard(0), validation would reject correct inputs.

This PR fixes both by pytree-flattening inputs in _compute_expected_inputs and forward(), and by extracting the actual solved placements from the solution dict via get_plain_input_and_grad_nodes. This also removes the duplicated default placement logic from _compute_expected_inputs, making it a pure shape calculator.

Other cleanups along the way:

• _extract_input_info now returns per-tensor devices (previously _make_input_fn hardcoded device="cuda")
• apply_placement takes sharding_placement as a required argument (every call site already passed it explicitly)

Authored with Claude.