Unit test avoids creating new contexts unless it needs specific configuration options

[rroelke]

This was at attempt at SC-63578. Performance results indicated that more vCPUs correlates negatively with performance. We speculated that this was because of the time required to spawn and join threads for each context. This branch reduces the number of contexts created by adding functions vanilla_context_c and vanilla_context_cpp which return an existing Context for unit tests which don't need any special configuration.

Ultimately this does not appear to be the cause of the negative correlation. time ./tiledb_unit shows that this branch runs in just under 21 minutes instead of just over 22 on a large EC2 instance with 96 vCPUs, which is only a 5% improvement. That's nothing to write home about, but the work is already done, so we might as well take it.

In addition to the vanilla_context_c and vanilla_context_cpp functions, there is refactoring of the VFSTestSetup class, splitting it into VFSTestContext and VFSTempDir.

The non-trivial changes live in test/support/src/{helpers,vfs_helpers}.{cc,h}.

---

TYPE: NO_HISTORY
DESC: Create fewer contexts in tiledb_unit

[teo-tsirpanis]

Hmm, on the one hand I'm not sure if we should merge it if the performance gains are not significant (for comparison, on a two-core machine the tests run one order of magnitude faster), but on the other hand this work will be useful when the C API gets fixed to make contexts thread-safe, but if we wait until then we risk bitrotting the PR.