optimize gpu.cu which leads to a ~4% speedup

[mrzetti]

This is created by me trying out gpt5.6 sol xhigh.
Further testing on more gpus and cuda versions would be nice :)

Changes

• transpose convolution tables from [256][6] to [6][256] to avoid stride-6 shared-memory bank conflicts
• reduce Perlin convolution kernels from 16 to 12 gradients by canonicalizing duplicate hashes:
  • 12 -> 0
  • 13 -> 9
  • 14 -> 1
  • 15 -> 11
• use uint8_t indices so modulo 256 occurs once instead of at every lookup
• use separate input/output buffers for filter_2_01b, fixing unsafe in-place compaction

Resource impact

On sm_89, filter_gradvecs_1 shared memory drops from 17,712 to 16,560 bytes per block. This allows six resident blocks within Ada's 100 KiB shared-memory limit instead of five.

Benchmark

RTX 4070, sm_89, bounded small biomes, fixed start seed 0:

Version	Throughput
Previous local baseline	~15.5 Gseeds/s
This change	~16.2 Gseeds/s

Fixed-seed stage counts remained identical during testing.

Related work

PR #24 also modifies the gradient-vector kernels, but takes a different approach by expanding and flattening the convolution buffers. This change instead reduces and transposes them, lowering filter_gradvecs_1 shared memory from approximately 30 KB in #24 to 16,560 bytes while preserving fixed-seed stage outputs.