Out-of-bounds read in `train_from_buffer` when sample sizes overflow

`ZDICT_trainFromBuffer` is never told the length of the samples buffer, so
`train_from_buffer` checked that the sample sizes add up to it. With
overflow checks off, as in release builds, that sum wrapped around: sizes
totalling 2^64 plus the buffer length passed the check, and the trainer read
past the end of the buffer. `zstd::dict::from_continuous` relies on this
check from zstd 0.4.10; older versions of zstd had their own, covered by a
separate advisory.

This requires the sample sizes to come from an untrusted source. The flaw
was corrected by summing the sizes with overflow checks.