Add ability to control PTX max registers and other per entry tuning parameters per Func.

[zvookin]

Per this PR:
#1667
being able to set PTX tuning parameters per Func would be useful. When (or if) the LLVM support to control .maxnregs lands, we should add a .ptx_tune(...) scheduling directive of some sort.

[abadams]

Pinging Ravi as an FYI @ravi-teja-mullapudi

[zvookin]

I think the llvm support now exists. How do we want to map this to scheduling directives?

[zvookin]

From the PR referenced above, it looks like @abadams suggested "f.cuda_tune(some_enum, some_value)". So this issue is now to design and implement the cuda_tune scheduling directive.

[abadams]

We've decided to leave this up to the driver, giving it the appropriate information (#9284)