Reduce heuristic grid size, int32 -> short

[Elucidation]

[REVERTED] Premise was wrong: dist grids are integers increasing to max dist

worst case example 200x200 grid max val is 40,000, so need int32 at least (int16 is -32 to 32k).

This does lead to #98 however

---

Default heuristic grid size is int32 for each entry. Since we're using values of in range -1-4 we don't need a 32-bit int, a signed short should be more than enough.

World Shape: (26, 24), 44 robots, 48 item zones, 24 stations
Building true heuristic
Built true heuristic grid in 275.42 ms
Dict with 116 keys,
Size: 296,032 bytes, one entry: 2496 dtype = int32
NxM = 26x24 = 624
Expected size ~NxMxZ = 26x24x116 = 72,384 values
For int32 = 4 bytes, so 4 bytes * 72,384 = 289,536 bytes
Measured 296032 vs 289536, diff =  102.24 %

[Elucidation]

Using np.short at 2 bytes, we are down from 290kb to 145kb, can we go smaller and single byte?

Building true heuristic
Built true heuristic grid in 579.28 ms
Dict with 116 keys,
Size: 151,264 bytes, one entry: 1248 dtype = int16
NxM = 26x24 = 624
Expected size ~NxMxZ = 26x24x116 = 72,384 values
For int16 = 2 bytes, so 2 bytes * 72,384 = 144,768 bytes
Measured 151264 vs 144768, diff =  104.49 %

[Elucidation]

Using np.int8 we are down to 1 byte

orld Shape: (26, 24), 44 robots, 48 item zones, 24 stations
Building true heuristic
Built true heuristic grid in 590.16 ms
Dict with 116 keys,
Size: 78,880 bytes, one entry: 624 dtype = int8
NxM = 26x24 = 624
Expected size ~NxMxZ = 26x24x116 = 72,384 values
For int8 = 1 bytes, so 1 bytes * 72,384 = 72,384 bytes
Measured 78880 vs 72384, diff =  108.97 %

And we've reached ~72kb from our previous 290kb, 4x savings

[Elucidation]

Strangely, using int8 seems to both slow down generation of the heuristic (by multiples!) and when running queries. Possibly due to numpy BLAS implementation

With np.int32

World Shape: (26, 24), 44 robots, 48 item zones, 24 stations
Building true heuristic
Built true heuristic grid in 197.99 ms
Dict with 116 keys,
Size: 296,032 bytes, one entry: 2496 dtype = int32
NxM = 26x24 = 624
Expected size ~NxMxZ = 26x24x116 = 72,384 values
For int32 = 4 bytes, so 4 bytes * 72,384 = 289,536 bytes
Measured 296032 vs 289536, diff =  102.24 %
---
Testing true heuristic grid on 1000000 queries
Did 1000000 queries into true heuristic grid in 278.76 ms

With np.int8

World Shape: (26, 24), 44 robots, 48 item zones, 24 stations
Building true heuristic
Built true heuristic grid in 569.73 ms
Dict with 116 keys,
Size: 78,880 bytes, one entry: 624 dtype = int8
NxM = 26x24 = 624
Expected size ~NxMxZ = 26x24x116 = 72,384 values
For int8 = 1 bytes, so 1 bytes * 72,384 = 72,384 bytes
Measured 78880 vs 72384, diff =  108.97 %
---
Testing true heuristic grid on 1000000 queries
Did 1000000 queries into true heuristic grid in 471.28 ms

[Elucidation]

Also realizing even with uint8 a 255 max distance path in a grid limits our grid size to a much smaller number than is allowed, so we can't use int8 anyway.

[Elucidation]

Also noticing google cloud instance cOS default python int is int64 (8 bytes), whereas on dev machine windows its int32 (4 bytes)