feat[gpu]: CUDA dynamic dispatch kernel

[0ax1]

What changes are included in this PR?

• Dynamic dispatch GPU kernel for fused decoding operations
• Support for source operations (right now only bitunpack) and scalar transforms (alp, for, zigzag)
• Documentation and usage guide

What is the rationale for this change?

Fusing kernels avoids intermediate memory copies to global memory by chaining operations together in a single dynamic dispatch kernel.

How is this change tested?

• Tests covering dynamic kernel dispatch with bitunpack, FoR, ALP

Are there any user-facing changes?

• New DynamicDispatchPlan API for composing decoding operations
• User guide in dynamic_dispatch.md

[0ax1]

benchmark on gh200:

bitunpack_for_alp/dynamic_dispatch_u32/100M
                        time:   [163.98 µs 165.58 µs 169.77 µs]
                        thrpt:  [2194.3 GiB/s 2249.8 GiB/s 2271.8 GiB/s]

2600GiB/s being the ref, running the bit-unpack kernel (6 bit) without dynamic dispatch in isolation.

[joseph-isaacs]

LG, whats the plan to hook this into the executor?