[WIP] Derive a scope's symbol table once per scope

[ThrudPrimrose]

symbols_defined_at walks every descriptor in the SDFG, and propagate_memlet derived it once per
memlet, so building a program cost O(scopes x descriptors) sympy work -- on a 2000-line generated
kernel, 8845 walks over 2082 descriptors, 81% of a 183 s parse. Callers propagating several memlets
through one scope node now derive it once, Data.free_symbols is memoized behind the property
setter that is the only way its sources can change, and the descriptor scan reads s.name instead
of printing the symbol.

tests/corpus/cloudsc, build_cloudsc_sdfg(simplify=False)   -- this branch vs main
  before  872.6 s      after  800.3 s      (6342 blocks, 2687 descriptors)

warpx_field_gather (hpcagent_bench, 2000 lines), to_sdfg(simplify=False)
  before  182.6 s      after   38.4 s      (7269 blocks, 2948 descriptors)

The second kernel is measured on the extended branch: it needs a data-dependent shape main's
frontend refuses, so main cannot parse it either way.

[ThrudPrimrose]

Holding this on the extended branch for now: the win is large on generated kernels and small on the cloudsc corpus, so it wants more evidence before landing on main.

[ThrudPrimrose]

Reopened -- keeping it as a draft. The change is running on the extended branch meanwhile, which is where the generated-kernel numbers come from.