flatten builds, use rattler-build

[minrk]

• switch to rattler-build
• flatten build matrix, so duplicate libdolfinx isn't built (fewer, longer jobs because the python matrix isn't parallelized anymore)
• windows still (temporarily) skipped

[minrk]

These shouldn't be required, but are for now to workaround prefix-dev/rattler-build#1317

[minrk]

With flatter builds, loads of time is saved because 80% of libdolfin builds are redundant (assuming a matrix of five Pythons). But the actual time-to-completion is still longer because of conda-forge's parallelism. But it does mean we are hogging a lot of resources, especially for the 40 arm/ppc builds, which take much longer than the rest, especially in the Python test phase.

Some summary of times:

	flat (no python matrix)	split (python matrix)	conda-build (python matrix)
builds	20	100	100
wall time	93 min	63 min	86 min
serialized CI time	18 hours	32 hours	47 hours
shortest build	18 min	6 min	14 min
longest build	93 min	41 min	45 min

So if everything's going smoothly, a lot of resources are saved by flattening builds like this. But if/when things start failing (e.g. the frequent intermittent network issues we've been seeing over the last week or so), the flatter builds significantly impact the turnaround time and cost of restarted builds. Notably, the shortest merged builds are only slightly slower than the shortest single-python builds with conda-build, so maybe this isn't so bad.

[minrk]

@conda-forge-admin please lint

[minrk]

I'll merge this one once conda-smithy releases the lint fixes

[conda-forge-admin]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/recipe.yaml) and found it was in an excellent condition.