Delete the six duplicated src/abi/*.idr modules and repoint ~15 references #815

What

Six Idris2 modules exist twice, with identical blob SHAs in both locations:

src/abi/hypatia-abi.ipkg sets sourcedir = "..", which resolves to src/,
so modules = Hypatia.ABI.Types compiles src/Hypatia/ABI/Types.idr. The
src/abi/*.idr copies are compiled by nothing.

They are not harmless. PR #811 found that
test/unified-api-adapter-contract_test.exs bound its golden source to the
dead src/abi/Types.idr — so the live ABI could drift and the guard would
stay green. That was instance 21 of the guard-asks-a-different-question trap,
and it was fixed in #811. The duplicates that made it possible are still here.

Blast radius — measured

~15 files name the dead path. Only the Elixir guard read it (now fixed);
the rest are prose and doc comments:

schema/README.adoc, ffi/zig/README.adoc, verification/PROOF-STATUS.adoc,
docs/operations/prover-wars-integration.adoc, ffi/connectors.json, doc
comments in ffi/zig/src/main.zig, ffi/zig/src/unified-api-adapter.zig,
clients/rust/hypatia-client/src/connector.rs, .../lib.rs, .../types.rs,
.../ffi.rs, verification/proofs/idris2/ConfidenceBounds.idr, a
data/verisim record, test/sensor_test.exs.

Acceptance criteria

• The six src/abi/*.idr files are deleted.
• idris2 --build src/abi/hypatia-abi.ipkg still succeeds — proving the
  ipkg was compiling the live tree all along, which is the claim this issue
  rests on.
• idris2 --build src/abi/hypatia-abi-gen.ipkg still succeeds and
  just abi-gen still produces a zero diff.
• Every one of the ~15 references is repointed to
  src/Hypatia/ABI/<Module>.idr. Verify by a repo-wide grep for
  src/abi/.*\.idr returning only the .ipkg files.
• grep -c before and after is recorded in the PR, so the denominator is
  stated rather than "all references updated" being asserted.
• mix test and .github/workflows/abi-codegen-drift.yml both stay green.