Dynamic memory management preset + updated wgpu buffer memory management

[nathanielsimard]

It's now more stable and much faster. The maximum memory usage is probably higher than before, but this is because it shows the maximum amount of memory used by the GPU, not the current amount. There is no deallocation, and we allocate in a way that is fragmentation-friendly for better speed, possibly at the cost of slightly higher maximum memory usage.

[mepatrick73]

Looks good to me! Maybe when we do chunk removal, we don't want to reset the cursor_chunk to 0, but for now I think it's fine.

[codecov]

Codecov Report

Attention: Patch coverage is 85.50725% with 20 lines in your changes missing coverage. Please review.

Project coverage is 85.64%. Comparing base (d696d74) to head (7c985a4).
Report is 4 commits behind head on main.

Files	Patch %	Lines
...ates/burn-compute/src/memory_management/dynamic.rs	83.52%	14 Missing ⚠️
...-compute/src/memory_management/memory_pool/base.rs	78.94%	4 Missing ⚠️
...-compute/src/memory_management/memory_pool/ring.rs	83.33%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1962      +/-   ##
==========================================
+ Coverage   85.23%   85.64%   +0.41%     
==========================================
  Files         810      810              
  Lines       99753   100154     +401     
==========================================
+ Hits        85022    85776     +754     
+ Misses      14731    14378     -353     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mosure]

it is unclear if a change in this PR introduces this behavior but after migrating from 71bd5efbfaf220b5058dbf4f963326e0b0d38527 to 69be99b802dc35d6eafbf49611776575f0ef906e, I see this assert panic:

thread 'main' panicked at .\burn-178c6829f420dae1\69be99b\crates\burn-compute\src\memory_management\memory_pool\small.rs:123:9:
assertion `left == right` failed
  left: 0
 right: 32

    fn allocate_slice(&self, handle_chunk: ChunkHandle, slice_size: usize) -> SmallSlice {
        let slice = self.create_slice(0, slice_size, handle_chunk.clone());

        let effective_size = slice.effective_size();
        assert_eq!(effective_size, self.buffer_storage_alignment_offset);

        slice
    }