Collapse quadratic forms to a single triangular solve

[jessegrabowski]

Computing b.T @ K^-1 @ b applies K^-1 with a forward and a back substitution, and then the surrounding dot contracts the result straight back to a scalar. Writing K = C C.T, the same quantity is ||C^-1 b||^2, so one substitution does. The rewrite catches cho_solve and also a hand-written pair of transposed triangular solves, whose factor doesn't have to come from a Cholesky; the square-root Kalman filter gets its factor out of a QR.

The first commit lifts the b_ndim=2 gate on psd_solve_to_chol_solve so a positive-definite solve with a vector right-hand side reaches the same chain instead of staying one opaque posv call.

The two benchmarks differ in whether the triangular factor is already available. In given_factor it is, and dropping one substitution roughly halves the time. In with_cholesky the factorization is redone on every call, so its O(n^3) cost dominates a single quadratic form and the saving shrinks to 6-8%, which is within one standard deviation except at n=3000.

------------------------------------ benchmark 'n=1000': 4 tests -------------------------------------
Name (time in us)                                         Median              StdDev            Rounds
------------------------------------------------------------------------------------------------------
test_quadratic_form_given_factor[1000-baseline]         281.9160 (2.08)      11.5286 (1.49)       3444
test_quadratic_form_given_factor[1000-rewritten]        135.7915 (1.0)        7.7555 (1.0)        7107
test_quadratic_form_with_cholesky[1000-baseline]      2,174.2499 (16.01)    117.2891 (15.12)       463
test_quadratic_form_with_cholesky[1000-rewritten]     2,044.7075 (15.06)    128.6777 (16.59)       509
------------------------------------------------------------------------------------------------------

----------------------------------- benchmark 'n=200': 4 tests ----------------------------------
Name (time in us)                                     Median             StdDev            Rounds
-------------------------------------------------------------------------------------------------
test_quadratic_form_given_factor[200-baseline]       15.3743 (1.85)      2.0301 (1.74)      46155
test_quadratic_form_given_factor[200-rewritten]       8.2916 (1.0)       1.1643 (1.0)       17532
test_quadratic_form_with_cholesky[200-baseline]      60.7083 (7.32)     18.1682 (15.60)      9088
test_quadratic_form_with_cholesky[200-rewritten]     56.1252 (6.77)      6.3358 (5.44)      15545
-------------------------------------------------------------------------------------------------

---------------------------------- benchmark 'n=3000': 4 tests ----------------------------------
Name (time in ms)                                      Median            StdDev            Rounds
-------------------------------------------------------------------------------------------------
test_quadratic_form_given_factor[3000-baseline]        3.0421 (1.90)     0.2504 (1.47)        329
test_quadratic_form_given_factor[3000-rewritten]       1.6029 (1.0)      0.1705 (1.0)         571
test_quadratic_form_with_cholesky[3000-baseline]      26.6957 (16.65)    0.5196 (3.05)         50
test_quadratic_form_with_cholesky[3000-rewritten]     24.6633 (15.39)    0.8209 (4.81)         50
-------------------------------------------------------------------------------------------------