parse: Implement hacky way to handle full range of small ints.

[pfalcon]

So, I proceeded to write testcases for varlen encoded small ints, and that's why I paid attention that parser has its own "small ints" which don't match normal "small ints". So, while my aim was to let people write a billion as a number in uPy, that still can't be done. OMG, let's work around it then, but see TODO below.

Parser has its own "small small ints", which unfortunately means that full
range of small ints are not available in lexical form. This issue is usually
masked by presence of long int, but without them, it's not possible to write
biggish numbers in a Python program.

To work that around, CONST_INT node is overloaded with special encoding for
full-range small ints, which eventually propagated to bytecode properly, as
LOAD_CONST_SMALL_INT.

TODO: Ideally, parser should not have its own "small int" type, and instead
all parts of interpreter should use single consistent type (even if that on
parser side requires more fat representation) - dealing with types fragmented
without real need is a chore.

[dpgeorge]

How about we fix it properly now. Proposal is to change the encoding of parse node leaves to following (for 16 bits, easily generalises to 32, 64 bits):

0000 0000 0000 0000  no node
xxxx xxxx xxxx xx00  pointer to mp_parse_node_struct_t (x != 0, will always be 4 byte aligned)
xxxx xxxx xxxx xxx1  small int, x being the value
xxxx xxxx xxxt tt10  other, t=type, x=value

[pfalcon]

Well, that's nice encoding scheme! I didn't think it would be easy to change tagging away from current 4-bit scheme, so my thinking revolved around having another buffer to store small ints, and put index # into mp_parse_node_t . But indeed, if it's not compound node, and not small int, then payload is either qstr id or token id, so tags for such nodes can be easily extended from 4 to 5 bits.

So, let me know if you would have time for this refactor, otherwise I can give it a try.

[dpgeorge]

Give it a shot!

[pfalcon]

Superseded by #314.