Refactor assemble

[connorjward]

This pull request:

• Lifts zeroing the tensor out of _assemble(). This means that get_scalar(), get_vector() and get_matrix() only need to return tensor, rather than tensor, zeros, lambda: tensor.
• Caches parloops on the form instead of constructing a monolithic assembly 'thunk'. This makes it straightforward to pass arguments into the parloops at execution time and will speed assemble() up significantly.

The changes will introduce a small amount of overhead where create_assembly_callable() is used because we do more logic every time the assembly callable is called. If this overhead is problematic we could also cache the callables that apply the boundary conditions.

What do people think? I know the code needs documenting and the caching does not yet work but I wanted to check before spending more time on this.

[connorjward]

Parloops are now cached on the form. I've done some quick performance checks and (in the pathological case of a very simple kernel and tiny mesh) assemble() is now about 4 times faster! Unfortunately, it has slowed down the create_assembly_callable() alternative by a factor of around 2 because we now do more work at every call. I have some ideas about how this could be reduced though.

[connorjward]

Pending tests and reviews I think this is ready to go.

[ksagiyam]

Looks good to me!

[ReubenHill]

@connorjward I would really appreciate you taking us through these changes and the rationale behind them just so we're all up to date. Maybe at the next firedrake meeting? (Whenever that happens!) Looks like you've done some really nice work here

[connorjward]

@connorjward I would really appreciate you taking us through these changes and the rationale behind them just so we're all up to date. Maybe at the next firedrake meeting? (Whenever that happens!) Looks like you've done some really nice work here

Sure, very happy to! I can summarise what I've done pretty straightforwardly though:

• We can now cache the parloops on the form (see _assemble_expr()). The current way I've implemented this is quite restrictive but this is likely to change with the other work that I'm doing.
• The logic is now a lot tidier (in my opinion).

[ReubenHill]

I've been annoying but, I hope, useful and gone through the docstrings to see if everything makes sense to me. See the various comments I've left for more.

[connorjward]

@wence- any chance you could review this at some point?