Add NDCG metrics

[exitflynn]

Adds NDCG metrics as a part of adding RecSys metrics in #2631

This PR builds on the work done in #3054 and #2632 (much thanks to ili0820 and kamalojasv181!)

The primary changes left at this stage as per my understanding would be around vectorising parts of the DCG computation as previously discussed here and here, and then docs.

I've considered different ways to address that, namely using torch.vmap (but I think that will be better to pursue once pytorch/pytorch#124423 is resolved); a batched vectorisation and a full one. Right now I think the batch-based vectorisation could be a good way to go since a full vectorisation approach could be a bit too complex (/overkill?) here, the main cause of complexity being different numbers of unique values across samples. Let me know what you think! I'm still in the early stages of working at this level and so open to feedback is all I am 🤠

[vfdev-5]

Thanks for the PR @exitflynn
Concerning the vectorization, can you please check how others implement this, for example here:

• https://github.com/catalyst-team/catalyst/blob/master/catalyst/metrics/functional/_ndcg.py
• https://github.com/zuoxingdong/recsys_metrics
• maybe, some other libs

The part to try to vectorize is _tie_averaged_dcg_batched. I do think vmap will ever support getitem operator.
What we can do first is to run some benchmarks on the same data between ignore_ties=True vs ignore_ties=False and see how slow _tie_averaged_dcg_batched implementation.
Next, we can try to improve it, maybe padding with zeros can help?
I haven't yet checked the details.

[vfdev-5]

@exitflynn if you have any questions, blockers, feel free to ask here or on our discord: https://pytorch-ignite.ai/chat