candle upgrade to 0.9.2alpha2

[michaelfeil]

• upgrades candle to latest rc
• I feel like ~5% overall improvements in performance.

https://github.com/michaelfeil/candle-extensions-fork -> Can be PR'ed and merged into huggingface's candle fork.

[michaelfeil]

There is some issue with the static-linking feature. It seems like some crates are requesting dynamic-linking by default. Once we are selecting static-linking inside this repo, but some external crates select dynamically-linking, we have a cudarc both dynamically and statically linked, which is not allowed.

[alvarobartt]

Hey @michaelfeil thanks a lot for this PR, as already mentioned we're directly working with candle to make sure that the linking strategy doesn't default to dynamic-linking for neither cudarc nor intel-mkl-src so that we can simply bump candle without having to rely on a fork to "remove" the default linking strategy. I'll ping you on the candle PR later today, but once that's done IMO we can bump candle here too, and eventually add support for Flash Attention 3 as per your PR at #785, thanks again and sorry for the inconvenience 🤗