Support `assign` argument in `LightningModule.load_from_checkpoint` method

Nobody has claimed this yet.

### Description & Motivation

Currently, `LightningModule.load_from_checkpoint` only allows to pass `strict` argument to `torch.nn.Module.load_state_dict` method. There is also a new `assign` argument (https://github.com/pytorch/pytorch/pull/102212) that can be used to efficiently load model parameters from a checkpoint (see discussion in https://dev-discuss.pytorch.org/t/state-of-model-creation-initialization-seralization-in-pytorch-core/1240). This is a FR to allow passing of `assign` argument to `load_state_dict` similar to the `strict` argument.

If this is something desired I can open a PR.

### Pitch

Implementation can be similar to the `strict` argument. One consideration is to make it backward-compatible (any suggestions on how to?).

### Alternatives

_No response_

### Additional context

_No response_

cc @lantiga @borda

Start at LightningModule.load_from_checkpoint and trace where it calls torch.nn.Module.load_state_dict. Follow the existing strict argument handling and determine how assign can be passed while preserving compatibility with older PyTorch versions. Done means callers can provide assign for checkpoint loading without breaking existing strict-only usage.

Written by the indexing model from the issue text.