fix out-of-range double-to-int UB in proto map key lookup

[sahvx655-wq]

The four ValueAsInt32/ValueAsInt64/ValueAsUInt32/ValueAsUInt64 helpers in parsed_map_field_value.cc coerce a double map key to the proto key's integer type, but they run the static_cast before checking the value is in range. Converting an out-of-range, negative (for an unsigned key), or non-finite double to an integer type is undefined behaviour, and a -fsanitize=float-cast-overflow build aborts with "1e+19 is outside the range of representable values of type 'long long'". It is reachable from any expression that indexes or tests membership on an integer-keyed proto map field with a double key, which CEL accepts ({1: 'x'}[1.0] is defined to work), for example msg.map_int64_int64[1e19] or the same value passed through dyn().

internal/number.h already gets this right for in-memory maps by range-checking the double before narrowing. The fix routes each double branch through cel::internal::Number, so the lossless-convertibility check runs first and AsInt/AsUint only see a value known to be representable. Behaviour for valid keys is unchanged and previously-undefined keys now report a miss. All four sibling helpers shared the defect, so all four are covered, with a regression test over the int32/int64/uint32/uint64 map fields.

[sahvx655-wq]

gentle ping

[jnthntatum]

/gcbrun

[sahvx655-wq]

The red check is the external builder sitting at action_required rather than a real failure. Its own note says it needs the SHA-pinned form, /gcbrun(84e77731f38b4d7d0509311d9e05de6031014283), to build the latest commit on the fork, so the bare /gcbrun didn't actually kick it off. Nothing in the diff is breaking it.

I rebuilt on this commit locally to be sure: //common:value compiles clean and the new FindDoubleKeyConversion case passes. The fix only reroutes the double branch through internal::Number so the lossless-range check runs before the narrowing cast, matching what the in-memory map path already does, so an out-of-range, negative, or non-finite key reports a miss instead of tripping float-cast-overflow. Glad to rebase or adjust anything if you need it.