more MPI memory management on large inputs

[sbailey]

This PR makes 3 updates to help with MPI memory management on especially large inputs like jura special-other healpix 27257 with 33749 targets (!):

• when broadcasting the large RESOLUTION HDUs, break into chunks to avoid 2 GB per-bcast limit
• when collecting results, break into chunks to avoid 2 GB per-send limit
  • this is a followup to PR #310 which worked up to ~20k targets, but not ~30k
• if the input communicator is larger than needed, split it into a sub communicator with just the ranks that will be use to avoid unnecessary bcasts to unused ranks

I verified on main/dark healpix 0 that the outputs are the same as main, whether the chunks are big or small. A final test on special-other healpix 27257 is still running; I will confirm that succeeded before merging, but I wanted to get this PR in while that was running.

[coveralls]

coverage: 38.62% (-0.1%) from 38.759% when pulling 517ca1c9198e52dc7fa8865265168bbf37f6b681 on bcast_mem into 30095434a0adc96698f0606cc35f7ce0a9a50b4f on main.

[coveralls]

coverage: 38.62% (-0.1%) from 38.759% when pulling 3ff0daf7105e6ba157bbe012d9f4ab3428bc9ea8 on bcast_mem into 30095434a0adc96698f0606cc35f7ce0a9a50b4f on main.

[sbailey]

Test job succeeded on 33749 targets in 25 minutes. After discussion with @dmargala I removed the unnecessary send/recv tag logic, since the blocking operations enforce the order anyway. Will merge and retag for use with Jura. This is backwards compatible with output of previous version.

[coveralls]

coverage: 38.425% (-0.3%) from 38.759% when pulling 3ff0daf7105e6ba157bbe012d9f4ab3428bc9ea8 on bcast_mem into 30095434a0adc96698f0606cc35f7ce0a9a50b4f on main.

[coveralls]

coverage: 38.509% (-0.3%) from 38.759% when pulling 3ff0daf7105e6ba157bbe012d9f4ab3428bc9ea8 on bcast_mem into 30095434a0adc96698f0606cc35f7ce0a9a50b4f on main.