Move the buffer copy logic outside operators

[anagainaru]

If an operator (direct or inverse) is not applied (due to a threshold or some other reason) the Operate or Decompress functions are now returning 0 and are not responsible with copying the data to the ADIOS2 internal buffers.

This allows ADIOS2 to use the appropriate function for copying the data (taking into consideration the memory space).

Operators that use this: Blosc, MGARD, MGARDPlus

Testing Adding test for MGARD with small GPU buffers.

Note: the original code inside operators was using memcpy so I am simply moving the logic outside (I am not checking for reverse endian)

Adding to #3465 to skip copy for blocks with 0 elements

[anagainaru]

@JieyangChen7 I use 1024 as the threshold for when to apply MGARD based on my testing on Summit but you might have a better idea of what value to use.

[JieyangChen7]

@anagainaru I think 1e6 might be a safer choice based on my experience.

[anagainaru]

@pnorbert I will merge this after I rebase it over release_29 and if it doesn't solve Axel's issue I will update his PR (#3465) with the fix. I think it make more sense like that compared to changing this PR further.

The base branch was changed.