Add HEVD solver, hermitian_tag, is_hermitian check and .H operator property

[jpbrodrick89]

This adds a new Hermitian eigenvalue solver HEVD which uses jnp.linalg.eigh under the hood. Don't worry this is not me moving to support general eigenvalue querying but merely stands as a more efficient version of SVD for Hermitian rank-deficient operators. There is little controversial about this PR except if there was a specific reason the .H property was not added initially and the branchy way of removing singular eigenvalues when tagged as rank deficient.

Changes made:

• Adds a new lineax.HEVD that performs Hermite eigenvalue decomposition on Hermitian operators
• HEVD has awareness of eigenvalue ordering when statically truncating singular eigenvalue based on `MaxRankTag``
• AutoLinearSolver(well_posed=False) dispatches to HEVD by default for Hermitian operators
• Adds is_hermitian check. To ensure custom user linear operators don't raise with AutoLinearSolver(well_posed=False) provides a default fallback implementation which queries symmetry and definiteness. All lineax operator receive dedicated registrations even if their implementation is identical to the fallback.
• is_symmetric will register True if is_hermitian and real.
• Adds hermitian_tag and wires it into all tag propagation (tags from checks, transpose tags, invert tags)
• Adds .H operator property which calls operator.conj().transpose() unless the operator is Hermitian, in which case it no-ops.
• Adopts .H in _linear_solve_jvp_rule to take advantage of no-op and also queries is_hermitian to skip conjugate transposing the state.
• Applies the optimisations first floated in jvp optimisations for pseudoinverse solvers #217 by noticing that a lot of rectangular solver have a "Gram partner" (SVD -> HEVD, Tall QR -> Cholesky, Tall Normal -> inner) as such we can can call linear_solve_p with the Gram partner when calculating the $(A^H A)^\dagger$ term.

I'm open to alternative names if there are strong opinions, options include EIGH, EigH, HermitianEigenValueSolver but I prefer the obvious analogue of HEVD with SVD.

@adconner this will affect #221 as we will need to cover Hermitian indefinite rank-deficient (I think I've added the old-style versions here). Note I'm definitely happy to get rid of the test that asserts iterative solves raise for singular matrices as I triggered a configuration where this failed.