Add support for `torchshard`

Nobody has claimed this yet.

## 🚀 Feature

[torchshard](https://github.com/KaiyuYue/torchshard) is a newly and simple library for model parallel. In some case, we need to split one very large layer (larger than one gpu) to multiple gpus, and `torchshard` seems a simple solution.

Here's one of my use-case:

1. I have a very large tensor (9M*1024), aka. `VeryLargeTensor`, which is the representation vectors of 9M examples pre-computed from their features.
2. Given a new example `e`, I use model `Model` to generate `e`'s representation `v`(1*1024).
3. I need to compute dot-product similarity between `v` and `VeryLargeTensor`, and get a 9M*1 score tensor. However, `VeryLargeTensor` is so large that it must be split into multiple gpus.
4. During training, only `Model` is updated. `VeryLargeTensor` is always freezed and never updated.
5. (Since `VeryLargeTensor` is a plug-in tensor and not updated, I don't want to register it as `Model`'s parameter.)6.
6. If I split `VeryLargeTensor` in to multiple gpus, I do not need DP or DDP.

I think `torchshard` is a simple solution for my use-case. Or, is there better solution in pytorch-lightning?
### Motivation

### Pitch

### Alternatives

### Additional context

______________________________________________________________________

cc @borda @akihironitta @justusschock

Start by reviewing the torchshard project linked in the issue and the repository's existing distributed training support. Clarify whether support means integration, documentation, or an alternative approach for sharding a frozen tensor across GPUs; the issue is complete only after that scope and an acceptance criterion are defined.

Written by the indexing model from the issue text.