Do not merge: working PR for NASA crew

[romanc]

Description

dace/math.h re-exposes a couple of math functions from cmath, e.g. std::sin(x), which has overloads for e.g. float and double. Some functions, e.g. asin(x), were not exposed in this way. This leads to precision issues when asin(x) is called with a float because in the generated code, asin will be mapped to the C version, which is only defined for doubles. There is thus an implicit cast happening of the argument and the computation is done in double precision. Worse, the return type is always a double, which will force the rest of the calculation to be up-casted to double precision if asin() is used in an expression.

TODO list for later

• make test case for math functions (e.g. compare against std::log10(x) in a cpp-tasklet)
• make test case for symbol replacement (double-replace with strings, replace symbol with non-default dtype)
• make test case for fix: stree -> sdfg, data descriptors of nested maps
• make test cases for memlets_in_ast() (subscript access, scalar access, indirect array access)
• add support for indirect array access in memlets_in_ast()
• once memlets_in_ast() is reliable, we can inline conditionals and loop condition/update (in stree as well as in gt4py bridge)
• refactor tn.IfScope / tn.ElifScope / tn.ElseScope to tn.ConditionalScope
• validation succeeds if conditional uses a scalar that isn't defined (anymore); see image below; might be fixed by considering scalar-memlets (to be re-tested)
• validation succeeds if mapped tasklets aren't fully connected; see image below

[alexnick83]

This looks fine. I would add a test that demonstrates the issue described in the PR (up/downcasting of data) and confirms that it is now resolved. I guess just one test for, e.g., asin would be enough.

[romanc]

Make sense, I'll add something tomorrow. I think I can build upon compare_numpy_output().

[romanc]

putting this back into a draft state because it is now our de-facto working branch and I've started to push more unrelated changes. I'll pull this apart later and make nice PRs with unrelated changes and tests.