[Packer] Updated the Next State Logic of Iterative Re-Packing

[AlexandreSinger]

The packer iteratively re-packs the netlist if it fails to create a dense enough packing. The way it does this can be thought of like states, where each state turns on a new set of packer options that can create a denser clustering.

The prior state transition logic for the iterative re-packing did not turn on enough options to make the clustering as dense as VTR can manage. This updates the state transition logic to create even more dense clusterings.

The next state logic is as follows:

• If there is an over-utilized block type, turn on unrelated clustering and balanced block type utilization if able.
• If there are overfilled floorplan regions, create more attraction groups than the prior iteration (or increase the pull). If all regions have attraction groups and the pull is strong, do not increase further.
• Finally, if all else fails, increase the target external pin utilization if any of the overfilled block types have target pin utilizations less than 1.0.
• If all prior fails, fail hard.

Found that for the Z1000 architecture, this increased the maximum LUT utilization from around 68% to around 86%.

Some data from the Z1000 architecture:

• Before this change, only the 4 smallest MCNC benchmarks could pack on the Z1000 architecture:

circuit	num LUTs	Lut Utilization
tseng	1046	51.1%
ex5p	1064	52.0%
apex4	1262	61.6%
misex3	1397	68.2%

• After this change, 6 more circuits were able to be packed on the Z1000 architecture:

circuit	num LUTs	Lut Utilization
dsip	1362	66.5%
diffeq	1494	73.0%
alu4	1522	74.3%
des	1591	77.7%
bigkey	1699	83.0%
seq	1750	85.5%

• Two MCNC circuits can still theoretically fit on the Z1000, but are still failing to be packed:

circuit	num LUTs	Lut Utilization
apex2	1878	91.7%
s298	1930	94.0%

Resolves #2729

[AlexandreSinger]

@petergrossmann21 FYI

[AlexandreSinger]

@vaughnbetz This PR should resolve your comments on the last PR. Let me know what you think.

[vaughnbetz]

LGTM -- thanks!