Fix COVER tiny training set underflow

[Rohithmatham12]

Fixes #4682.

COVER_ctx_init() validates totalSamplesSize before computing suffixSize, but the subtraction uses trainingSamplesSize. With splitPoint < 1.0, a small training split can be smaller than MAX(d, sizeof(U64)) even when the total sample size passes validation. That lets the unsigned subtraction underflow and can lead to an oversized allocation.

This adds the missing trainingSamplesSize precondition before computing the partial suffix array size, and adds a regression for ZDICT_optimizeTrainFromBuffer_cover() with ten one-byte samples, auto-selected d/k, and splitPoint = 0.5. The API now returns srcSize_wrong instead of attempting the underflowing allocation.

Test plan:

• make -j4 fuzzer
• ./fuzzer -i1 -s1 -v
• git diff --check