Cleaner HEXL NTT integration

[fboemer]

• Avoids a slowdown on first-time NTT computation due to computing NTT objects upon first use (fixes HEXL slowing down some operations in Microsoft SEAL benchmarks intel/hexl#7, see also Initial Intel HEXL integration #312 (comment))

  • This helps reduce confusion when using a small number of iterations for benchmarking, as in SEAL's examples/benchmark code. E.g. running example 7 / option 1 (BFV with default degrees) yields
    • Before this PR: (N=8192) Average batch: ~148 microseconds (on first run; faster on second run)
    • After this PR: (N=8192) Average batch: ~57 microseconds (on each run)

• Also reduces the memory consumption: before this PR, the ntt_cache_ was stored in each compilation unit separately due to using a static variable in the header file. Running ./build/bin/sealbench yields:

  • Before this PR : [ 2590 MB] Total allocation from the memory pool
  • After this PR: [ 2388 MB] Total allocation from the memory pool

• Fixes the HEXL commit hash, which was mistakenly pointing to the main branch instead of the 1.1.0 branch (see https://github.com/intel/hexl/tags).

[WeiDaiWD]

The static fix is quite nice!

I noticed that intel_seal_ext.h and intel_seal_ext.cpp are guarded by SEAL_USE_INTEL_HEXL. In a future major release, I will add these two files to the CMake build system only if SEAL_USE_INTEL_HEXL is true. This will avoid installing "empty" headers. Similarly to ztools.h and ztools.cpp.