Fix FSDP2 crash with ignored_params on torch < 2.7.0 (#3924)
The `ignored_params` keyword argument was unconditionally passed to
`fully_shard()`, but this parameter is only supported in torch >= 2.7.0.
On older versions (e.g. 2.6.0), this caused a TypeError when using
FSDP2 with LoRA adapters or any configuration that triggers the
fsdp2_prepare_model code path.
Move the `ignored_params` kwarg out of the base fsdp2_kwargs dict and
only add it conditionally when torch >= 2.7.0 and ignored_modules is
actually configured.
Fixes #3923
