fix(topology): a cube's units are contiguous in ABSOLUTE_POS

[ThierryCantin-Demers]

CubeDim::new always returns a 2D cube, and ABSOLUTE_POS is linearized grid-major, so its y term strides the whole grid. For a cube of (32, 8, 1) on a grid of (2112, 1, 1) it comes out as ty*(gridDim.x*32) + bx*32 + tx, and the eight planes of one cube land on eight streams 1.03 MiB apart instead of one contiguous run.

Each warp is still fully coalesced, so the generated code is unchanged. What changes is how many concurrent streams reach DRAM, and stores are far more sensitive to that than loads.

This linearizes it as cube_pos() * CUBE_DIM + UNIT_POS in all four backends.

Measured

RTX 4070 Ti SUPER, the throughput example, three runs each:

	before	after
memory write 512 MiB	581.4 GB/s	596.4	+2.6%
memory copy 1 GiB	592.1 GB/s	597.3	+0.9%
memory read 512 MiB	638.9 GB/s	638.4	unchanged

RTX 2060, same example, three runs each, main against this branch on the same day. This card has a 3 MiB L2 against the 4070's 48 MiB, and pays far more for scattered streams:

	before	after
memory write 512 MiB	255.7 to 256.9 GB/s	284.9 to 285.9	+11%, 76% of bus to 85%
memory copy 1 GiB	251.3 to 252.3 GB/s	269.2 to 269.7	+7%
memory read 512 MiB	305.3 to 307.8 GB/s	319.3	+4%

The after column is within two points of what the same card reports under Vulkan, which was the only backend on it that reached those rates before.

Copy is the out[i] = f(in[i]) shape most elementwise kernels have, so about 1% is the realistic figure for those. Read is unchanged, which is expected: only the store side is stream sensitive.

cubek kernels

cubek main pinned at this branch's base against this branch, RTX 4070 Ti SUPER, same day. Rows are one strategy on one shape; a negative change is faster.

bench	rows	median	range	runs per side
reduce	132	0.0%	-0.7% to +2.6%	1
gemm, tensor-core strategies on 4096³ f16/f32 and a skinny shape	12	+0.0%	-0.1% to +0.4%	2
gemv	22	+0.2%	-13.4% to +3.1%	1
conv2d	2	-0.7%	-8.1% to -0.7%	1
unary	3	+1.4%	-5.3% to +1.8%	3

The two large moves are gains: the autotuned gemm entry on the vector-matrix shapes, -13%, and the AlexNet-shaped conv2d, -8%. Unary at vec4 is 1.5% slower in all three rounds, the one row pointing the other way; at vec8 it is 5% faster in two of three.

Compatibility

ABSOLUTE_POS no longer linearizes (ABSOLUTE_POS_X, Y, Z). Both forms are bijections onto the same range, so a kernel computing f(ABSOLUTE_POS) covers identical indices and only which unit gets which index changes. The axes themselves are unchanged, and nothing outside their definition sites pairs them with ABSOLUTE_POS.

Tests

The existing topology test writes value i to index i and compares against 0..n, which holds under any bijection, so it cannot see this. Added one that writes CUBE_POS to each unit's ABSOLUTE_POS slot and asserts contiguous runs per cube. It fails on main and passes here, on CUDA and on Vulkan.

suite	result
cubecl-cuda runtime tests, RTX 4070 Ti SUPER	933 pass, 17 fail: the same 17 as the base commit (16 complex math, test_cmma_manual on sm_89)
cubecl-cpu runtime tests	766 pass
cubecl-wgpu runtime tests, Vulkan	743 pass
burn main against this branch	builds
cubek main against this branch	builds; kernels benchmarked above

[louisfd]