That optimization was too complex and caused some weird traces when
debugging. The more common case «if ... then break» was kept.

Instead of an explicit value (field 'b'), true and false use different
tag variants. This avoids reading an extra field and results in more
direct code. (Most code that uses booleans needs to distinguish between
true and false anyway.)

- Several details in 'lcode.c'
- A few more tests for code generation
- Bug in assert in 'lcode.c' ("=" x "==")
- Comments in 'lopcodes.h' and 'ltable.c'

The difference in performance between immediate operands and K operands
does not seem to justify all those extra opcodes. We only keep OP_ADDI,
due to its ubiquity and because the difference is a little more relevant.
(Later, OP_SUBI will be implemented by OP_ADDI, negating the constant.)

Plus, this commit removes useless 'tm' parameters in 'op_*' macros.

In arithmetic/bitwise operators, the call to metamethods is made
in a separate opcode following the main one. (The main
opcode skips this next one when the operation succeeds.) This
change reduces slightly the size of the binary and the complexity
of the arithmetic/bitwise opcodes. It also simplfies the treatment
of errors and yeld/resume in these operations, as there are much
fewer cases to consider. (Only OP_MMBIN/OP_MMBINI/OP_MMBINK,
instead of all variants of all arithmetic/bitwise operators.)

Attributes changed to posfixed ('x <const>', instead of '<const> x'),
and "toclose" renamed to "close". Posfixed attributes seem to make it
clearer that it applies to only one variable when there are multiple
variables.

Constants directly assigned to other constants were not propagating:
For instance, in
  local <const> k1 = 10
  local <const> k2 = k1
'k2' were not treated as a compile-time constant.

String literal expressions have their own kind VKSTR, instead of the
generic VK. This allows strings to "cross" functions without entering
their constant tables (e.g., if they are used only by some nested
function).