fix(train.py): mfu estimation to respect CPU-GPU sync point

[JasonLiJT]

Previously, the mfu timing measurement was taken before the CPU-GPU sync point at every iter. The resulting running_mfu:

• would converge correctly when log_interval = 1.
• could converge to > 100% when log_interval > 1.
  • This could create the illusion that bumping log_interval speeds up training (it usually does not).

See diagrams below.

log_interval = 1

log_interval = 2

Note that t3 - t2 is discarded. Only t2 - t1 and t4 - t3 etc contribute to running_mfu.