This is a complete rewrite of the code in hash.c

Move from a chained hash table implementation to open addressing with
Robin Hood probing. This allows to increase the maximum fill factor and
further reduce the growth factor, saving considerable amounts of memory
without sacrificing performance.

To make this work, hash values are now cached in the table entry
also avoiding many key comparisons.

Tables are created lazily with a smaller minimum size.

Insertion functions now report an error if growing the table resulted in
a memory allocation failure.

Some string comparisons were optimized to call directly into libc
instead of using the xmlstring API.

The length of inserted keys is computed along with the hash improving
allocation performance.

Bounds checking was made more robust.

In dictionary-based mode, unneeded interning of strings is avoided.