[WIP] tridiagonal_eigh

[paolot-gc]

This PR covers the mods required to implement the computation of eigenvectors and eigenvalues of a (symmetrical) tridiagonal matrix, such as the one emerging from a reduction to Hessenberg form.

[paolot-gc]

At commit c0e27ff, ipu_tridiag_solve() runs a Vertex implementing Thomas algorithm.
One can run python examples/tridiag_solver_example.py 1472 1472 to solve 1472 symmetrical tridiagonal linear systems with random coefficients, each with 1472 unknowns, concurrently.
Profiling indicates this takes a total of 176,743 cycles. Memory usage was 46.7MiB (memory excluding gaps).
Note that at present, the algorithm is implemented as a Vertex, therefore using just one of the 6 threads.

[paolot-gc]

With commit 85b4ebd, I added Alex's code for the computation of eigenvalues of a tridiagonal matrix and added an example in which eigenvectors of a real symmetric matrix are computed using the inverse iteration method.
The chart below shows the execution trace for the computation of the 1472 eigenvalues and eigenvectors of a 1472x1472 symmetric matrix. Only two inverse iterations were performed.
The vast majority of the 46.7M cycles are taken by the Hessenberg tridiagonalization.

There was a small but non-negligible "reconstruction" error.

Compiling module jit_real_symmetric_eigh.0:
[##################################################] 100% Compilation Finished [Elapsed: 00:04:44.4]
ic| np.max(np.abs(mat @ x.T - eig * x.T)): 0.00011253357

Memory usage is reported below. Note the huge component due to Internal Exchange Message Buffers (276MiB) and Internal Exchange Code (115MiB).