Skip breakpoints on read-only bytecode pages instead of aborting

[tmikov]

The debugger installs breakpoints by overwriting an opcode in the bytecode stream with Debugger, which requires the page to be writable. The patcher uses mprotect to make the page RW; on failure it called hermes_fatal and aborted the process.

If the bytecode lives in a read-only segment (e.g. statically linked into the binary as a const array, ending up in __DATA_CONST / __TEXT_CONST on macOS), mprotect is rejected by the OS (EACCES under hardened runtime) and the abort takes down the whole process. This is reachable in practice because step / restoration / on-load breakpoints install at offset 0 of every CodeBlock about to execute, so stepping into such a CodeBlock from anywhere kills the runtime.

Make the patch best-effort:

• CodeBlock::installBreakpointAtOffset and the static makeWritable helper now return bool. On failure no state is modified.
• Debugger::installBreakpoint returns BreakpointLocation* (nullptr on failure), rolling back the partial breakpointLocations_ entry it just inserted.
• doSetNonUserBreakpoint and setOnLoadBreakpoint silently skip when installBreakpoint fails. Step / restoration / on-load breakpoints are best-effort hints; the only loss is that stepping cannot pause at the entry of a CodeBlock whose page cannot be patched.
• setUserBreakpoint now returns bool so callers can propagate failure to their clients (e.g. CDP could surface a proper error back to DevTools instead of aborting).
• The four call sites that re-install a previously-installed breakpoint (the page is known writable since the original install succeeded) assert the result.

[meta-codesync]

@tmikov has imported this pull request. If you are a Meta employee, you can view this in D105190367.

[facebook-github-tools]

@tmikov has updated the pull request. You must reimport the pull request before landing.