Add short/long benchmark suites with kernel-specific shapes

[debashishc]

Summary

Define explicit benchmark suites with kernel-specific shape grids, including a long suite for large-N performance.

Motivation / Use Case

We want deterministic, comparable benchmark runs across kernels and GPUs.

Proposed Solution

• Add short and long suites in bench/suites.yaml.
• Each kernel declares its own shape list:
  • reduce_sum and softmax_online use dim=-1 shapes from Issues 1 and 2.
  • copy_transpose uses shapes from Issue 3.
• Document usage in README/DEVELOPMENT.

Scope Alignment

v0.1 scope (Weeks 0-2)

Alternatives Considered

Keep only a single smoke suite.

Additional Context

Long suite is optional and benchmark-only.