Reduce number of nodes used to store ValueRecords

[cmyr]

In doing some simple profiling I noticed that we are allocating a lot of nodes when parsing variable metrics in value records. We're going to need to reduce the amount of nesting here. As illustration, when compiling the variable features for Google Sans, we encounter the following node counts:

VariableMetricNode         59676 
LocationValueNode          477408
LocationSpecNode           477408
LocationSpecItemNode       1432224
AxisLocationNode           1432224

If we can start out by removing AxisLocation and LocationSpecItem that would reduce our allocations by 70%.