Sparse mode performance and SparseHist input dispatch

[bendavid]

Adds support for sparse histogram input in the TensorWriter.

Significant performance optimizations for sparse mode in
both the TensorWriter and in the Fitter.

Performance optimizations should also give a factor of ~2
improvement for dense mode in the Fitter for large models.

Note that this depends on WMass/wums#25

Two related groups of commits:

Group A — TensorWriter sparse dispatch (7 commits)

• add option to treat input systematic histograms as difference with respect to nominal
• add test for sparse mode
• Support scipy sparse array inputs in TensorWriter and add as_difference option
• Add multi-systematic dispatch in add_systematic and use wums.SparseHist
• Add external likelihood term (gradient + hessian) support
• Add efficient SparseHist multi-systematic dispatch in TensorWriter
• Speed up TensorWriter for large multi-systematic SparseHist workloads

Group B — Sparse fast path performance (5 commits)

Up to ~20× HVP speedup on the jpsi calibration tensor (76800 bins,
108334 params, 62M-nnz logk, 329M-nnz external sparse Hessian):
HVP 6380 → 320 ms, loss+grad 3010 → 160 ms.

• inputdata, parsing: prep for sparse fast path with CSR matvec
  — canonicalize sparse index ordering at load time, pre-build a
  CSRSparseMatrix view of logk, add --hvpMethod and
  --noJitCompile CLI options.

• fitter: dynamic loss/grad/HVP wrappers with jit_compile +
  hvpMethod — replace class-level @tf.function decorators with
  instance-level wrappers built dynamically in _make_tf_functions,
  so jit and HVP autodiff mode can be controlled per-fit. Note that
  fwdrev HVP is intentionally never jit-compiled because
  tf.autodiff.ForwardAccumulator does not propagate JVPs through
  XLA-compiled subgraphs.

• fitter: sparse fast path uses CSR matmul, no dense
  [nbins, nproc] — reformulate the sparse branch of
  _compute_yields_noBBB to use tf_sparse_csr.matmul for the
  inner contraction logk @ theta (~8× faster per call than the
  equivalent gather + segment_sum) and never materialize the dense
  [nbins, nproc] grid in the NLL/grad/HVP path. Also forces
  jit_compile=False in sparse mode (CSR matmul has no XLA kernel)
  and falls back to revrev when fwdrev is requested in sparse mode.

• fitter: external sparse Hessian via CSR matmul — switch the
  external sparse-Hessian likelihood term to use CSR matmul. The
  registered gradient of sm.matmul is itself a single
  sm.matmul, so reverse-over-reverse autodiff no longer
  rematerializes a 2D gather/scatter chain in the second-order
  tape. On the jpsi 329M-nnz prefit Hessian this was the dominant
  HVP cost.

• rabbit_fit, setup.sh: enable XLA multi-threaded Eigen on CPU —
  set XLA_FLAGS=--xla_cpu_multi_thread_eigen=true so XLA's CPU
  emitter uses Eigen's multi-threaded routines for the dense
  matmuls jit_compile=True generates. ~1.3× speedup on dense
  large-model HVP/loss+grad on a many-core system. Set both in
  setup.sh (for sourced shells) and at the very top of
  bin/rabbit_fit.py before any TF import (for direct invocation).