[Proposal] Change zstd compression levels mapping

[LRFLEW]

I noticed that bcachefs is mapping its compression levels (1-15) to zstd's compression levels (1-22) using floor(level * 3 / 2). This PR proposes to change this mapping to be direct, and use only zstd compression levels 1-15 (see later comments about the new mapping). My reasoning for this change is as follows:

• This mapping, requiring integer truncation, results in an uneven mapping of compression levels. Increasing the bcachefs compression level by one will increase the underlying compression level by either one or two depending on the starting level. Using a direct mapping of compression levels fixes this.
• bcachefs uses a default compression level of 3, which seems to map to zstd's default compression level of 3. However, the remapping of compression levels means that a bcachefs compression level of 3 is actually a zstd compression level of 4, which is more than the default for zstd. Using a direct mapping will bring the default compression levels of bcachefs and zstd in line with each other.
• Referring to zstd's compression levels as being "from 1 to 22" is a bit disingenuous, as it ignores the seven "negative" compression levels (accessed in the command line tool with --fast=N). Having bcachefs ignore some of the higher compression levels is not a significant deal, considering that it's already ignoring some of the lower compression levels.
• Similarly, with the command-line tool zstd, the compression levels 20-22 are not accessible without an additional command-line argument (--ultra) to tell it you really want to use compression levels that high. Counting those higher compression levels as being equivalent to the others can be misleading, and excluding them shouldn't be a significant change.
• btrfs's zstd compression also has 15 compression levels, but they map their compression levels to zstd's compression levels 1-15 directly. Changing bcachefs's compression level mapping to match btrfs means that existing benchmarks for btrfs compression levels will be more accurate for bcachefs.

If keeping access to these higher compression levels is a priority, then perhaps a non-linear mapping would be better suited. One option would be to map the lower levels directly to the zstd levels, and then have the higher levels skip every other compression level (clamped to <=22). This would provide a compromise between providing more low compression levels and maintaining access to the max compression level, all while addressing the above issues. Let me know if you'd rather this solution, and I can update the PR with it.

[boomshroom]

Since background_compression happens, well, in the background, it's not as important for it to be fast and it makes sense to let the system use spare background cycles getting the maximum possible compression for very infrequently accessed files. As such, being able to access level 22 for this purpose would be nice and I'd appreciate not losing it.

If we wanted a piecewise-linear mapping with a slope of 1 at (0,0) and a slope of 2 at (15,22), then they would meet at exactly (8,8), giving one extra point to the lower range.

[LRFLEW]

I noticed that piecewise-linear mapping while experimenting with it, and I do think that's a good option. The only issue I have with it is that levels 20-22 might want to be considered separate.

Levels >= 20, labeled --ultra, should be used with caution, as they require more memory. (source)

Because of that, I think level 19, the maximum non-ultra level, should be included in the mapping. As such, I would propose using the mapping points (0,0) and (14,19), giving an intersection point of (9,9). Then the top level 15 would be mapped to the top ultra level 22. The only possible issue I see is that it adds additional complexity to the mapping.