Householder reduction to Hessenberg form

[paolot-gc]

This branch implements reduction to Hessenberg form using Householder reflections. It makes use of the vertices developed for the QR decomposition.

[balancap]

Up to you on how you want to do it, but I am happy if you want to merge the initial implementation of Hessenberg, and then work on the for_iloop version in a new PR/branch. Just need a small unit test before merging (like checking vs Numpy).

Personally a big fan of short PR quickly merged :)

[balancap]

Btw, for the CI to pass, you'll need to run the pre-commit: https://github.com/graphcore-research/tessellate-ipu/blob/main/docs/development.md#development
Should be mostly formatting + a bit of flake8 checks

[paolot-gc]

I made the changes as per the comments above.
In the accuracy test (i.e. comparison against the scipy implementation), I had to lower the precision (decimal=4) compared to what was set for the QR decomposition (decimal=5).

[paolot-gc]

Just to document the effects of the commit above:
This execution trace below (before the addition of ipu_data_barrier()) shows that the Q matrix updates (top tiles) occur after the R matrix updates (bottom tiles)

After moving the Q matrix update and adding the `ipu_data_barrier()', the updates occur in parallel

[balancap]

Thanks @paolot-gc , looking great 👍