Some drawcall submission bug fixes

[hrydgard]

Fix some bugs that my recent optimizations introduced.

Also, bring over a couple of minor changes from #18328 , in order to be able to rebase that one to shrink it.

[fp64]

Offtopic here, but any reason you use inverse scrambler from https://github.com/skeeto/hash-prospector (lowbias32_r)? It seems slightly more expensive than plain lowbias32, and significantly worse as an integer hash. Some numbers:

Function	Linf	RMS	PractRand
MurmurHash3's finalizer	4044016	566904.4	Fails at 16KB
lowbias32	2023971	372660.4	Fails at 32KB
lowbias32_r	168685748	9288956.4	Fails at 16KB
TheIronBorn's version	1211488	231074.2	Fails at 64KB
My version	836260	121867.6	Fails at 1GB
triple32	167788	44857.8	Fails at 1GB ('mildly suspicious' at 8KB)

Linf and RMS are avalanche scores (lower is better), PractRand measure its quality as an RNG.
Not sure how much this quality difference matters in practice.
Speed-wise, MurmurHash3's finalizer, lowbias32, and the TheIronBorn's version are identical. My version is slightly slower (should be about the same as lowbias32_r), and triple32 slightly slower still.

[hrydgard]

No reason in particular, in fact I plan to redo all the hashing here at some point.. That's good input.