fix: make BiFManager builtin-index hash independent of the C++ standard library

[TsaiGaggery]

Problem

BiFManagerCommon::getHash() uses std::hash<std::string>, whose result is
implementation-defined and differs across C++ standard libraries. BiFManager-bin
bakes the hashes of builtin names into the generated OCLBiFImpl.h as
case <hash>ULL: labels at build time, and BiFManagerHandler::GetDepList
recomputes getHash() for each referenced function at run time to find its section.

When the host BiFManager-bin and the libigc runtime are linked against
different standard libraries -- the usual situation when cross-compiling IGC for
Android (a libstdc++ host tool generating the header while libigc is built with
libc++) -- the generated case labels and the runtime lookups disagree for every
name. The switch always falls through to {-1}, findAllBuiltins() collects
nothing, LinkBiF() links no built-in bodies, and every __spirv_* reference is
reported undefined reference to ... -> backend compiler failure
(CL_BUILD_PROGRAM_FAILURE). Native single-stdlib builds are unaffected, which is why
this only surfaces in heterogeneous-stdlib / cross builds.

Fix

Use a fixed 64-bit FNV-1a hash so the generated labels and the runtime lookups always
agree regardless of the C++ standard library. getHash() is defined once in
BiFManagerCommon and compiled into both the host tool and the runtime.

Validation

Reproduced cross-compiling IGC (LLVM 17.0.6) for Android x86_64: host BiFManager-bin
linked against libstdc++, libigc.so against libc++. Before this patch, ocloc compile
of any kernel using a builtin failed with
undefined reference to _Z33__spirv_BuiltInGlobalInvocationIdi. Binary proof: the
libstdc++ hash of that symbol appeared as an immediate inside libigc.so while the
libc++ value the runtime actually computes appeared 0 times. After this patch the
FNV-1a value matches on both sides, findAllBuiltins collects the builtins, ocloc
reports Build succeeded., and OpenVINO GPU inference runs end to end.