small tensor parallelism optimization

[Alex-GZ]

Just two small changes in tensor parallelism:

• Removed an unnecessary (in my opinion) cross gpu barrier in tp_broadcast, so the GPU that had finished the transfer earlier can start its work sooner. Maybe instead, there should be a synchronization done before the CPU->GPU transfer, in the 'src_dev>=0' branch.
• Rewrote the barrier to make it cost 2n-2 event wait calls instead of n(n-1), while preserving its function.

This PR had improved generation T/s in my setup by about 20% in the zone of CUDA API queueing bottleneck (at <20k context, 123B 4.5bpw model, 6 mostly different GPU, CPU with rather slow single-threaded performance and Linux) The impact may be a bit higher on Windows, probably.

[Ph0rk0z]

Tested it, and appears to work. Splitting 123b over 4x3090 I go from high 16/17t/s to 19.x t/s. Went back and forth a few times. Almost as good as splitting over 3 only. Should try that next and see if I get a speedup too.

[Ph0rk0z]

Some more tests. 4gpu, there is gain, 3 gpu there is about a .5t/s drop.

Can you check on your setup for odd/even gpu? Maybe there is a way to use both ways depending on device count.