Tests: store PgoConfig instead of pre-computed PgoData

The previous commit shipped `compile_exe` with an ugly
`make_pgo_profile` closure that captured a pre-computed
`Option<ExecutionProfile>` and just handed it back when called. That
asymmetry was a smell — the production paths (CLI, openvm-eth) hand
the pipeline a `PgoConfig` carrying serialized inputs, and the
closure deserializes + runs the guest to *compute* the profile.

This commit aligns the tests with the production shape:

- `GuestTestConfig.pgo_data: PgoData` becomes `pgo_config: PgoConfig`.
- A test-local static `make_pgo_profile` deserializes a `StdIn` from
  `inputs` and calls `execution_profile_from_guest` — same contract
  as the CLI's version.
- `pgo_config_from_stdin(pgo, max_cols, &stdin)` and
  `none_pgo_config()` helpers replace the ad-hoc
  `PgoData::{Cell,Instruction}(profile, …)` constructions.
- `compile_exe` collapses to a 3-liner: build the `StagedPipeline`,
  apply `with_select_defaults`, call `setup`. No more
  `Option<ExecutionProfile>` capture dance.
- Test call sites drop the pre-computed
  `execution_profile_from_guest(&guest, stdin.clone())` lines
  entirely (the pipeline now computes the profile internally on the
  cache-miss path).
- `prove_simple` / `prove_mock` / `prove_recursion` /
  `compile_and_prove` take `PgoConfig` directly.

Author: georgwiese