Additional hardening of index load path

[scsiguy]

Summary:
Address input validation gaps in the load path:

#: 1
FPE (divide-by-zero)
Location: ProductQuantizer::set_derived_values()
Root Cause: M=0 causes d % M to divide by zero
Fix: Validate M > 0 in read_ProductQuantizer and set_derived_values()
────────────────────────────────────────
#: 2
OOB vector access
Location: AdditiveQuantizer::set_derived_values()
Root Cause: nbits.size() != M causes access to nbits[i] beyond bounds
Fix: Validate nbits.size() == M in read_AdditiveQuantizer
────────────────────────────────────────
#: 3
OOB vector access
Location: ResidualQuantizer old format
Root Cause: Same as #2 but via read_ResidualQuantizer_old path
Fix: Validate nbits.size() == M in read_ResidualQuantizer_old
────────────────────────────────────────
#: 4
Integer overflow
Location: ProductQuantizer::set_derived_values()
Root Cause: d * ksub overflows when nbits is large
Fix: Use mul_no_overflow() for centroids.resize(), moved nbits > 24 check
before code_size computation
────────────────────────────────────────
#: 5
FPE (divide-by-zero) + stack overflow
Location: IndexLattice constructor
Root Cause: nsq=0 causes d/nsq divide-by-zero; garbage nsq causes stack
overflow in ZnSphereCodec
Fix: Validate nsq > 0, d > 0, d % nsq == 0, r2 >= 0 before construction
Files Modified

Reviewed By: mdouze

Differential Revision: D93486058

[meta-codesync]

@scsiguy has exported this pull request. If you are a Meta employee, you can view the originating Diff in D93486058.

[meta-codesync]

This pull request has been merged in d260887.