Add tuning notes for Hopper 16-bit dense GEMM

[hamuzhan]

Add tuning notes for Hopper 16-bit dense GEMM

Summary

Adds a "Practical tuning notes for Hopper 16-bit dense GEMM" subsection to
media/docs/cpp/efficient_gemm.md, at the end of the existing "Hopper Warp
Specialization" section.

The notes cover four points for selecting FP16/BF16 dense GEMM kernels on SM90,
based on profiling on H100 (CUDA 13.2) with the CUTLASS profiler and Nsight
Compute:

1. Accumulator register budget per schedule. The per-warp-group accumulator
   formula for warpspecialized, ping-pong, and cooperative schedules, and the
   local-memory spilling that happens for per-warp-group tiles larger than
   64x256. This matches the pruning in the companion PR.
2. Small-M (skinny/decode) problems. Why wide default tiles leave SMs idle at
   M <= 128, which tile/cluster/schedule shapes recover throughput (measured up
   to +44% at M=16 and up to +74% at M=128, FP16 and BF16), and the
   CUTLASS_LIBRARY_INSTANTIATION_LEVEL setting to generate them.
3. Rasterization and swizzle for memory-heavy shapes. Measured +8-10% from
   --raster_order=along_m --swizzle_size=2 on a DRAM-bound shape (median of
   interleaved runs), and a note that along_n costs about 25% on very-large-N
   shapes.
4. Sustained-clock effects. A measurement note for near-peak kernels under the
   board power limit.

Motivation

These are common performance issues for users selecting SM90 kernels by hand
rather than through autotuning. The existing section describes the kernel
designs but not the selection limits, so the notes go next to that description.

No code changes. Documentation only.