This avoids creating an HBytes for each and every code word. Instead, the
code words are collected into blocks behind each clear code and translated
together into a single HBytes per block.

This saves us from allocating and freeing the HBytes that were stored in
the table. It should also save memory since it essentially shares common
prefixes between codes.

The only remaining call to malloc() is the one allocating the global
context object itself.

Remember that HBytes itself just wraps a pointer and a size, so
this does not significantly enlarge the struct, but it saves a whole
bunch of allocation.

No need for it to be part of the exposed interface.

Commit 970f23cf already removed the use of h_butnot(), so there is no
need anymore for act_output to handle code = 257 (eod).

This includes the global context variable and all semantic actions and
validations. Besides being good practice, this makes the "LZW" in their
names unnecessary.

The only difference between the codeword and the litspec rules was that
the latter validated that code < 258. This has become redundant because
they were only still used for eod and clear both of which have their
own specific validation for the code value.

Thus the litspec rules and their validations can go.

This frees up the more generic name.

This makes LZW_literal redundant and removes the need to use h_butnot()
to detect eod.

This makes test/valid/lzw.pdf report decoder failure. Is that file actually
valid?

After the previous commit, we no longer need to know the last seen code.
The only remaining use was the test whether we have already assigned a
code (after clear). We can just as well detect that by inspecting the
number of defined codes.

This changes the logic of act_LZW_codeword such that it creates a new
table entry after processing each code word, even though it does not know
the last character, yet. We know that we will discover the last character 
on the next round, before we need it for any output.

In return we can remove all the fumbling around with prev_string.

A tiny gripe remains in the fact that HBytes declares its token member
const, so technically we are forbidden from filling in the last character
after the fact. But also technically, we can sledgehammer-cast the const
away, thanks.

Also slightly extends coverage of the defensive asserts and exposes a bug
(test/valid/lzw.pdf crashes) that I think must have been there before:
It seems that we never validate that code words are actually in the
defined range!?

This saves a tiny bit of code dup in updating ctx->old and building the
return value.

The only thing missing from act_LZW_codeword is to skip the table update
on the first code after a clear. The rest of the relevant code path is
virtually identical.

This logically matches the H_ALLOC in act_LZW_literal.

NB: We can drop the multiplication by sizeof(uint8_t) because the
latter is guaranteed to be 1. If uint8_t exists, CHAR_BIT must equal 8.

Fix recent instigator crashes

Closes #25, #31, #35, #36, and #37

See merge request !35

This removes the requirement to have at least one code word after a
clear code and fixes test/valid/lzw-clear2.pdf.

This "fixes" the assertion failure on test/valid/lzw-clear2.pdf but
exposes another issue: Codes 256 (clear) and 257 (eod) are actually
allowed to follow a clear code, yet our grammar says otherwise.

Includes a failing test. The decoder currently mishandles the code word
following a clear code. It expects a literal (codes 0-255) but accepts
the range 0-257. This appears to be the root cause of issues #26 and #37.

Fixes test/valid/lzw-clear.pdf

This makes the fall-through handling of TW_Tqq (reusing the TW_Tj
path) match the struct layout.

Fixes #36.

Fix several erroneous uses of dictentry/resolve. In particular, it
is fine to call resolve() on NULL (it is a no-op in that case), but
its result must be tested. It returns NULL, for instance, if a
referenced object does not exist.

The correct idiom is therefore

    val = dictentry(dict, "key");
    val = resolve(aux, val);

or

    val = resolve(aux, dictentry(dict, "key"));

depending on taste. I have preferred the former (two-line) variant.

The idiom should be encapsulated in a function. There are also
several occurances of dictentry() in the code that are not followed
by resolve() but probably should be.

Fixes #31, #35.

Drive-by fix a forgotten TODO item. ;)

This is a drive-by revert of a useless change in everyone's favorite commit,
6e5955c4 ("Most of the code folded in"). The line in question is a structure
assignment, which is in C99 and behaves exactly as one would expect.