Replacing uniform with uniform_philox , and incrementing PRNG state

[kiya00]

This PR is one step towards recomputing dropout in backward, the steps are:

• Add the prim operators to query/update the default CUDA RNG state PR
• Fix If saved_for_backward returns NumberProxy, the value is taken from compile time, not runtime #231
• Add transformation to replacing uniform with uniform_philox (this PR)
• Make sure the rematerialization works properly for Dropout

Ref #114

[kiya00]

TODO:
fix #231
update the offset according to nvfuser properly

[IvanYashchuk]

Let's split replace_uniform into a separate PR.
Primitives for getting and setting the state can be tested without replace_uniform and having #231 fixed is not required for these primitives.

[kiya00]

Hi @mruberry @jjsjann123 @IvanYashchuk , I modified it according our design review discussion, could you take a look?

[mruberry]

I'll let @IvanYashchuk do the final review. Thanks for the design review, @kiya00!

[IvanYashchuk]

merged 9 commits into rng_update_prims from uniform_rng

We had a manual stack here 😃

[t-vi]

Yeah, I'll learn to read...

[kiya00]

sorry for the trouble, I should get that draft status on