Parallelize detensorizing token IDs during tree search

[EricMichaelSmith]

Patch description
Currently, during any kind of tree search, the token ID tensor will be detensorized one item at a time, which is expensive. This PR rewrites that to detensorize the whole batch at once.

Performance on 1000 generations (-mf zoo:blender/blender_90M/model -t blended_skill_talk -ne 1000), on an otherwise unoccupied devfair:

• Original code, trial 1: median elapsed time 783 ms, mean 797 ms
• Original code, trial 2: median 770 ms, mean 785 ms
• This PR, trial 1: median 660 ms, mean 671 ms
• This PR, trial 2: median 656 ms, mean 667 ms

Testing steps
CI checks

[stephenroller]

so good.

[stephenroller]

do we need to port this change to torchscript?

[EricMichaelSmith]

do we need to port this change to torchscript?

@stephenroller Hmm, doesn't look like it, because torchscript export currently only supports greedy search