examples: add configurable MoE expert count and rows-per-expert to example 12

[knowrohit]

Feature

examples/12_xe20_moe_gemm_cute_interface gains two options:

• --experts=<int> — number of experts per layer (default 32, unchanged)
• --rows_per_expert=<int> — give every expert this many rows, replacing the
  built-in measured table (default 0 = use the built-in table, unchanged)

It also prints the effective configuration, so a result is self-describing:

MoE fill configuration
  Experts          : 32
  Layers           : 24
  Rows/expert      : built-in measured table
  Mean rows/expert : 250.824

Follow-up to #855.

Use Case

The example previously fixed num_experts at 32 and drew per-expert row counts
from a fixed 24x32 in-source table, so the GFLOPS it prints could only be
obtained at one fill unless you edited the source.

That matters because at fixed --n/--k the printed rate is dominated by fill.
On Arc Pro B70 (Battlemage G31), 32 experts, --n=2880 --k=2880 --num_layers=24 --verify=0, median over 24 layers of the N=5760, K=2880 leg:

rows/expert	TFLOP/s (median)	kernel ms (median)
built-in table	82.1	3.2510
uniform 250	112.7	2.3550
uniform 120	59.5	2.1419
uniform 70	35.7	2.0821
uniform 30	14.7	2.1663

From uniform 30 to uniform 250 the FLOP count rises 8.3x and the reported rate
rises 7.7x, while kernel time rises 1.09x — the leg is bound by streaming expert
weights at these fills, so the rate largely reports sum_e M_e. Routed MoE
inference operates near the low end of that table (top_k over tens to hundreds
of experts), which is exactly the region that previously required a source edit
to reach.

API

Additive only. Both options are parsed with the existing cutlass::CommandLine
idiom next to --n/--k/--num_layers/--verify, and validated with the same
error = true pattern used by examples/10_bmg_grouped_gemm_mixed_dtype:

• --experts must be positive
• --rows_per_expert must be non-negative
• --experts above 32 requires --rows_per_expert, since the built-in table
  only has 32 columns; the diagnostic says so explicitly

main() gained the standard if (options.error) { ... return -1; } block that
sibling examples already use.

The per-layer row array became a std::vector<int> because the expert count is
now a runtime value; the literal table is untouched and is now a
static const int [max_layers][kDefaultTableExperts] that seeds the vector when
--rows_per_expert is not supplied.

Example

# built-in table (current default, byte-for-byte the previous behaviour)
./12_xe20_moe_gemm_cute_interface --n=2880 --k=2880 --num_layers=24 --verify=0

# sweep fill at fixed shape and expert count
for r in 30 70 120 250; do
  ./12_xe20_moe_gemm_cute_interface --n=2880 --k=2880 --num_layers=24 \
      --verify=0 --rows_per_expert=$r
done

# a wider expert count, which requires an explicit uniform fill
./12_xe20_moe_gemm_cute_interface --n=2048 --k=3072 --experts=85 --rows_per_expert=30

Testing

Built and run on Arc Pro B70 (Battlemage G31), driver 1.15.39122+12, NEO
26.27.39122.12, oneAPI 2026.1, JIT spir64, SYCL_UR_USE_LEVEL_ZERO_V2=0.

• Builds clean via cmake --build build-sycl --target 12_xe20_moe_gemm_cute_interface.
• Default behaviour preserved. With no new flags the example reports
  Mean rows/expert : 250.824. I independently brace-parsed the 24x32 table
  literal out of the source (768 entries, every inner group verified length 32)
  and its mean is 250.8 — so the vector path reproduces the table exactly. The
  CMake test arguments --n=2880 --k=2880 --num_layers=24 are unaffected and
  need no change.
• Validation exercised. --experts=85 without --rows_per_expert prints the
  diagnostic and aborts with -1:

  ERROR: --experts=85 exceeds the 32 experts covered by the built-in row table.
         Pass --rows_per_expert=<int> to supply a uniform fill instead.
  Aborting execution.
  

• Sweep exercised. The five-fill table above, plus --experts=85 --rows_per_expert={30,120,230} at --n=2048 --k=3072, all run to completion.

Not tested: PVC, and verify=1 at --experts > 32 (the correctness reference
path was not touched, but I only exercised it at the default expert count).

ToDo

None that I'm aware of. If you'd rather have the flag spelled --fill, or want
the CMake test arguments extended to cover a second fill point in CI, say so and
I'll adjust.