Validate RV32I prebuilt count metadata

[thc1006]

rv32i-prebuilt.sh treats count.txt as optional and strips all non-digit characters before comparing it. As a result, a release without the file can pass, and a value such as 1x2 is interpreted as 12.

Require count.txt to contain exactly one positive decimal integer. Return 77 when the metadata is missing, malformed, or does not match the number of conformance ELF files. Add offline cases covering each condition.

make verify-prebuilt passes.
make check-all passes.

---

Summary by cubic

Enforces strict validation of count.txt in RV32I prebuilts to prevent truncated or miscounted suites.
rv32i-prebuilt.sh now requires exactly one positive decimal via rv32i-count.sh and compares as strings; missing, malformed, or mismatched counts exit 77, offline metadata cases and live smoke use the same helper, and the full pass-case runs only when count.txt matches the ELFs.

^{Written for commit 2d9a7d0. Summary will update on new commits.}

[jserv]

Reviewed the metadata parser, the guard ordering, the stubs, and the portability. The change is correct and well covered. The strict parser resolves every fixture plus the un-fixtured edges (empty, trailing newline versus a real second line, CRLF, whitespace, huge decimal), requiring count.txt does not regress a legitimate release (a demos-only release skips at the conformance == 0 guard first), and the set -e command-substitution idiom is portable in dash. The switch from -ne to a string compare also avoids the old arithmetic error on a malformed count.

[jserv]

Thorough fixture set. The parser also correctly rejects a few inputs not pinned here: CRLF (1\r\n), space padded ( 1\n or 1 \n), and it accepts an arbitrarily large decimal as a string. Fixtures for those would guard the behavior against a future parser change.

[thc1006]

Thank u Prof. for pointing out the remaining parser edges. I add fixtures for CRLF and leading/trailing whitespace, along with a direct parser check that preserves an arbitrarily large decimal verbatim. The direct check distinguish a valid-but-mismatched large count from a malformed count, since both would otherwise make the end-to-end consumer return 77.

[jserv]

Thank @thc1006 for contributing!