Fix a number of bugs in SyntaxTree and SyntaxTreeCode

[ChrisJefferson]

I believe this fixes #5117, but it would be good if @zickgraf could have a look.

Note that this changes some previously generated trees (see the diff), but I believe they were always wrong, it's just that the integer 1 was being interpreted as an Expr (where it is 0), but the integer 2 interpreted as an Expr caused a crash.

[fingolfin]

Ahhh, that makes sense. Nice catch!

[zickgraf]

Thanks a lot, works perfectly! And it confirms my expectation that a stack of EXPR_ELMS_LIST_LEV always terminates with an EXPR_ELMS_LIST, which means my solution in #5116 is actually correct :D

While reading the documentation again, I noticed that this level thing also exists for the usual list access (EXPR_ELM_LIST_LEV). Trying to use this gives a segfault which I guess has an analogous fix:

SYNTAX_TREE( function ( x ) return [ [ [ [ 1 ] ] ] ]{[ 1 ]}[1]{[1]}[1]; end )

Note: This is in no way important for me, I'm just trying to find further ways to produce crashes :D

[fingolfin]

Thanks @zickgraf , good catch. Perhaps @ChrisJefferson will also fix that, too :-)

[ChrisJefferson]

This PR grew slightly out of control while fixing some other SYNTAX_TREE issues, I'm happy to pull it apart.

Now it contains:

(outside of SYNTAX_TREE) while tracking down a crash I added asserts to SIZE_EXPR. This revealed for chars we make bags of size uchar(1), which mean we shouldn't the read and write Exprs (which are size 8) from them. As we already round allocations up to word size this wasn't actually doing any harm in practice, but let's fix it anyway.

I also (this is the slightly horrible code) generalised the SYNTAX_TREE code to support variadics in the middle of an object.

[ChrisJefferson]

Just to point out (in particular if @zickgraf is using this), this PR changes how matrix indexing is done, to make it (correctly) handle 2d matrices. Also (last push, as I'm sat in the airport), I decided to add an assert which makes sure all of an expr is used in SyntaxTree. This showed up a bug in handling EXPR_REC_TILDE. On deep investigation, it turns out this doesn't need to be handled differently (as we don't need to worry here where the tildes bind to, just preserve their presence).

Lots of changes were made after my review

[zickgraf]

Thanks for the effort! I have run all of my tests using this PR and there are no regressions :-)