Improve compile time (simple makefile changes)

[ademeure]

This improves compile times by:

1. Using nvidia-smi instead of __nvcc_device_query (the latter was wasting >1 second per build on a 1xH100 system!)
2. "--threads=0" allowss NVCC to use multithreading for some things (seems to improve compile times by ~5% locally.)
3. Making -O3 configurable with the FORCE_NVCC_O parameter (O0 improves compile times by >30% but significantly hurts run times; -O1 seems like a good compromise when iterating on some code.)

(1) and (2) reduce compile time with cuDNN from ~9.2s to ~8.1s.
(3) with O2/O1/O0 reduces it further to ~7.7s / ~6.5s / ~5.4s.

[karpathy]

nice!

[karpathy]

surprised about both 1 and 2...