Kronecker linalg rewrites

[jessegrabowski]

Adds several rewrites related to kronecker products that avoid materializing the full big matrices:

• solve_of_kron (solvers.py) — push solve through a Kronecker product:

  solve(kron(A, B), b)     ->  vec(solve(B, solve(A, unvec(b)).T).T)
  solve(kron(A, B).T, b)   ->  same, with A->A.T, B->B.T   (transpose peel)
  

• dot_of_kron (products.py) — push matmul through a Kronecker product (dual identity):

  kron(A, B) @ X     ->  reshape(A @ reshape(X, (m,p,k)) @ B.T, (m*p, k))
  kron(A, B).T @ X   ->  same, with A->A.T, B->B.T
  

• solve_of_kron_plus_diag_noise (solvers.py):

  solve(kron(K1, K2) + sigma**2*I, b)
      ->  Q @ ((Q.T @ b) / d)
          where  Q = kron(Q1, Q2),  d = kron_diag(d1, d2) + sigma**2
  

• det_of_kron_plus_diag_noise (summary.py):

  det(kron(K1, K2) + sigma**2*I)
      ->  prod(kron_diag(d1, d2) + sigma**2)
  

The prupose is to eliminate completely the KroneckerNormal distribution in PyMC, users will just get the best possible logp if they put a kronecker covariance matrix into a normal. solve and dot of kron are generally useful, but the ones with diagonal noise are pretty specialized, so I had them track KroneckerProduct itself, rather than Solve or Dot.