Add benchmark harness and tensor GEMM benchmark

[mmshad]

Summary

Transforms the repo from a collection of one-off SLURM scripts into a parameterized, architecture-aware benchmarking harness and adds a new tensor-core saturation benchmark (FP16/FP8).

Changes

Shared Infrastructure (common/)

File	Purpose
config.sh	Centralized SLURM/container defaults (partition, account, SIF path, modules). All values are env-var-overridable.
detect_arch.sh	GPU architecture detection via nvidia-smi compute capability. Gates precision support (FP8 on Hopper, FP4 blocked until Blackwell).
parse_results.py	Auto-detects benchmark type from .out files and produces Markdown summary tables for all 5 benchmark types.

Parameterized Launchers

Each benchmark (p1-p5) now has a single run_*.sh launcher that:

• Accepts configuration via environment variables (NODES, GPUS_PER_NODE, N, SLOPPY_TYPE, etc.)
• Auto-resubmits with proper --partition, --account, --constraint via exec sbatch --export=ALL if not already inside a SLURM job
• Replaces the need to maintain separate scripts per GPU/precision configuration

Benchmark-specific features:

• HPL -- Generates HPL.dat inline; supports sweep mode (HPL_SWEEP="92160:1:1 190464:2:2")
• HPL-MxP -- Maps friendly precision names (SLOPPY_TYPE=FP8) to numeric flags; auto-generates GPU affinity; gates FP4 by arch + container version
• HPCG -- Parameterized problem size and runtime
• STREAM -- Parameterized dtype (FP32/FP64) and array size

New: Tensor-Core GEMM Saturation Benchmark (p5_tensor_gemm/)

A cuBLASLt-based microbenchmark that measures raw FP16 and FP8 tensor-core throughput via dense GEMM. This answers: "How close can my GPU get to its theoretical peak tensor-core TFLOPS?"

• tensor_gemm_bench.cu -- ~440 lines, supports FP16 and FP8 (E4M3) on Hopper, FP4 gated to Blackwell
• run_tensor_gemm.sh -- SLURM launcher with smoke, saturation, transformer, custom modes
• 4 example scripts in examples/
• Auto-detects H100 vs H200, reports % of theoretical peak

Measured results (single GPU, 1980 MHz):

GPU	FP16 TFLOPS	% Peak	FP8 TFLOPS	% Peak
H100 SXM	928.4	93.8%	1654.8	83.6%
H200 SXM	885.5	89.5%	1595.7	80.6%

README Updates

• Added p5 section with theoretical peaks and measured results
• Added benchmark taxonomy table (Linpack-style vs tensor-core saturation vs memory/system)
• Added common infrastructure documentation
• Fixed GPU count errors in p1 and p2 READMEs ("2N x 4GPUs" = 8 GPUs total, not 4)
• Added sloppy-type precision table to p2

What's NOT Changed

• All original scripts in runs/ directories are untouched
• Existing benchmark output files preserved as-is

Testing

All benchmarks verified on H200 node (holygpu8a10302, 4x NVIDIA H200 SXM, 1980 MHz):

Test	Result
common/config.sh -- config loads, container path resolves	PASS
common/detect_arch.sh -- detects SM 90 / Hopper	PASS
p1 HPL -- N=8192, residual PASSED	PASS
p2 HPL-MxP FP16 -- N=8192, residual PASSED	PASS
p3 HPCG -- VALID result	PASS
p4 STREAM -- FP32 and FP64 bandwidth measured	PASS
p5 Tensor GEMM -- FP16 and FP8 saturation sweeps	PASS
FP4 gating -- correctly refused on Hopper	PASS