API changes to constructing DenseSkOps and populating their buffers, set next_state on construction, add web documentation, more stat tests

[rileyjmurray]

This PR is huge. I'm going to merge it even though there are outstanding documentation TODOs, as indicated in Issue #107.

Changes in RandBLAS proper

Major new identifiers

• SketchingDistribution and SketchingOperator C++20 concepts.
• isometry_scale_factor (overloaded for DenseSkOp and SparseSkOp), returns the scale that should canonically be applied to the full sketching operator assuming its entries are sampled in RandBLAS' standard way (uniform $\sqrt{3}[-1,1]$ or Gaussian(0, 1) for dense, or length-vec_nnz vectors of iid Rademachers for sparse).
• dense::compute_next_state and sparse::compute_next_state, for computing the appropriate value of S.next_state in O(1) time. This enables defining sequences of sketching operators with the same key without having to explicitly generate the operators.

Minor new identifiers

• For dense sketching operators, submatrix_as_blackbox accepts a sketching operator, row and column counts, and row and column offsets, and returns a DenseSkOp with a BlackBox distribution whose entries match those of the appropriate submatrix of the given sketching operator.
• safe_int_product (in RandBLAS/base.hh) checks for overflows. This is useful when computing potentially-large index offsets for implicitly defined matrices.
• util::omatcopy for general stride changes
• util::overwrite_triangle, for preparing a triangular matrix to be treated as a general matrix in gemm or trmm.
• A new value MajorAxis::Unknown in the MajorAxis enum, for use with BlackBox distribution families.
• An overload of repeated_fisher_yates that skips writes to vals and idxs_minor and that has more intuitive argument names.

Other changes

• The default scalar distribution for DenseSkOp objects is now Gaussian instead of uniform.
• The default scaling of uniform sketching operators is $[-\sqrt{3},\sqrt{3}]$, since this choice provides for unit variance.
• rewrite of dense::fill_dense_submat_impl, needed in order to efficiently implement dense::compute_next_state.
• In SparseSkOp, the typename RNG_t has been removed, a typename state_t has been added (equal to RNGState<RNG>) and the typename T_t typename has been renamed to scalar_t. Similar changes have been made to DenseSkOp for consistency across the two APIs.
• Changed the verbose overload offill_dense for writing an implicit submatrix to a given buffer. This function now requires a layout parameter as its first argument. It performs an out-of-place change of layout if the requested layout is different from the natural layout implied by the distribution object.
• Moved several function definitions for DenseSkOp and SparseSkOp into the struct definitions, rather than only declaring signatures in the struct definition and implementing the functions outside the struct.

Changes in documentation

• Added sample_indices_iid_uniform, sample_indicies_iid, and weights_to_cdf to a new page in the API reference.
• Added a section to the tutorial on updating sketches.

Changes in tests

New / significantly changed tests

• test_continuous.cc: Kolmogorov-Smirnov tests for distributional correctness of uniform [-sqrt(3),sqrt(3)] and Gaussian sampling. This includes a helper function for exactly computing the KS test statistic for distributions whose CDF is continuous, which has some surprising subtleties.
• test_r123.cc: a significant clean up of (and slight expansion of) test_r123_kat.cc.
• test_distortion.cc: statistical tests for subspace embedding distortion of dense sketching operators. This required a bunch of new infrastructure, described below.

handrolled_lapack.hh implements algorithms that would generally be considered "LAPACK-level," although the algorithms won't be found in LAPACK itself.

• cholesky (sequential and blocked)
• cholesky qr (with a flag to run a second pass)
• QR based on block classic Gram-Schmidt, where blocks are handled by cholesky qr2.
• QR with a two-pass version of blocked classic Gram-Schmidt. (Not equivalent to modified Gram-Schmidt.)
• A function to check if the outer-most Gershgorin discs for a positive definite matrix have converged to within some specified relative radius.
• cholesky iteration to compute all eigenvalues of a positive definite matrix (with termination criteria based only on extremal eigenvalues). This is actually prohibitively slow except when the matrix is very small. But it works, and it's tested, so might as well keep it around.
• power method to compute the extremal eigenvalues of a positive definite matrix. The number of iterations is chosen based on a well-known convergence result for the power method in the absence of a spectral gap. The smallest eigenvalue is obtained by Cholesky decomposing the matrix, explicitly forming its inverse, and then running the standard power method. This leads to faster performance for matrix dimensions and tolerances that we need in statistical tests.