[BSchilperoort]

This PR makes the model up to 16x faster on my GPU (AMD 7800 XT). The context was discussed in #138

The main performance improvements stem from:

• Keeping leaf classes as running totals, instead of computing a cumulative sum every time step
• Defer condition checking to end-of-day, instead of mid-step so they can be computed with only a single host-device sync (Carbon-balance, partitioning, latitude and vernalisation checks currently queue a device-side flag)
• Stop traitlets comparing every assigned tensor (Every state assignment ran bool(old == new). On a multi-element tensor this launches a comparison kernel and then fails silently; on one-element tensors it forces a sync)
• Move sets of operations into a single function, and compile these functions (e.g. in root_dynamics move computation of calc_rates into a separate compilable function)

In addition, the tests are sped up a bit too.

Disclaimer: the modifications were fully written by an LLM, and I did not go over every line of code. Tests were basically untouched in 5972458 (only 2 vars renamed), and all tests still seem to pass.

I am not in a hurry to merge this, and can wait for #137 to be merged first.