Remove thread local buffers

[rklaehn]

The thread local buffers are primarily for the case where you are decompressing something really small. For something larger, allocating is not an issue.

So just replace the entire stuff with this:

• if the input data is sufficiently small, try to decompress to a stack allocated fixed size buffer
• if not, or if this fails, just allocate the max size once and discard it again when done. Maybe there is some estimation of decompressed size in zstd_safe.