AP_Scripting: reduce Lua state memory usage

[tpwrules]

A few re-engineerings to make the state setup more efficient and on-demand. Please see the commits for details.

There are no visible changes to scripts except for ones that play games reading _ENV, which none in the repo do. There might be a slight downside of increased risk of fragmentation and out of memory problems at inopportune moments, but I think the savings are worth it.

Tested some scripts including my REPL in SITL and on Cube Orange.

Total savings on Cube Orange:

• 104 bytes of statically allocated RAM
• 14K of Lua heap memory (reduced by 2/3)
• ~800 bytes per running script
• 216 bytes of flash

[IamPete1]

Nice. Looks like this applies the same optimization we did for singletons to globals so we can keep them in flash rather than memory.

I think the only thing we need to check is how fast it is before vs after. I did do a little testing when we moved the singletons here:
#18249 (comment)

In that case it did approximately double the time taken to call a method. My concern here would be that we pay that on every call. Looks like ahrs is at the end of the list, so it will be the worst case. Its not clear to me if we have to do the search for each call or just the first? There is a DISABLE_INTERRUPTS_FOR_SCRIPT_RUN define that makes the timing more reliable.

What happens if we overload a singleton? EG ahrs = "foo"

[tpwrules]

Like the commit explains, the lookup result is stored in the global table so there should only be an additional cost the first time a particular singleton is accessed.

I will look into timing it to see if I can notice a difference. There might be a small new cost each call since the name has to fail to be found in the script environment before Lua looks in the global table. If this turns out to be unacceptable we can store the result in the environment at the cost of a bit more RAM usage.

If a script overrides a singleton name, then it will be successful, but will only affect that script as each script has its own environment separate from the global table. If the script deletes the override then it can access the singleton properly again.

[tpwrules]

Did some timing with the following script:

function update()
  local start_t = micros()
  local lt = 0
  for i = 0, 2000 do
    -- Location userdata creation function at the bottom of the table serach, so
    -- should be a worst case test for the lookup
    -- lt = Location -- Reference
    -- lt = Location() -- Instantiate
  end
  local end_t = micros()

  gcs:send_text(0, "took: " .. tostring(end_t-start_t) .. "us val:"
    .. tostring(lt))

  return update, 1000
end
return update()

And got some interesting results (times in microseconds):

Commit	Reference	Instantiate
master	2064	33562
Global table	3112	27279
Per-script table	1472	26062

Turns out that extra lookup costs more than I thought. It is probably faster overall in both cases in the instantiate case because the garbage collector has to do less work, so this should be a decent speed improvement overall.

Based on these results I am changing this PR to put the looked up entries into the script environment table instead of the global table. There is going to be a slight downside in that scripts now won't share userdatas, so if two scripts use the same singleton it will be ~32 more bytes than before. But that concern is probably overthinking it.