ILP not applied to variable loop length

Playing around with sparse / variable loop length codes, and came across this one

```
import loopy as lp
import numpy as np
lp.set_caching_enabled(False)

k = lp.make_kernel([
"{ [i] : 0 <= i < m }",
"{ [j] : 0 <= j < length }"],
"""
for i
<> rowstart = rowstarts[i]
<> rowend = rowstarts[i]
<> length = rowend - rowstart
y[i] = sum(j, values[rowstart+j] * x[colindices[rowstart + j]])
end
""")
k = lp.add_and_infer_dtypes(k, {'values,x':np.float64, 'rowstarts,colindices':k.index_dtype})
k = lp.split_iname(k, 'i', 2, inner_tag='ilp')
print(lp.generate_code(k)[0])
```

Generates:

```
#define lid(N) ((int) get_local_id(N))
#define gid(N) ((int) get_group_id(N))
#if __OPENCL_C_VERSION__ < 120
#pragma OPENCL EXTENSION cl_khr_fp64: enable
#endif
#define int_floor_div_pos_b(a,b) ( ( (a) - ( ((a)<0) ? ((b)-1) : 0 ) ) / (b) )

__kernel void __attribute__ ((reqd_work_group_size(1, 1, 1))) loopy_kernel(__global int const *restrict colindices, int const m, __global int const *restrict rowstarts, __global double const *restrict values, __global double const *restrict x, __global double *restrict y)
{
double acc_j[2];
**int length[2]**;
int rowend[2];
int rowstart[2];

for (int i_outer = 0; i_outer <= -1 + int_floor_div_pos_b(1 + m, 2); ++i_outer)
{
acc_j[0] = 0.0;
rowend[0] = rowstarts[2 * i_outer];
rowstart[0] = rowstarts[2 * i_outer];
**length[0]** = rowend[0] + -1 * rowstart[0];
if (-2 + -2 * i_outer + m >= 0)
{
acc_j[1] = 0.0;
rowend[1] = rowstarts[2 * i_outer + 1];
rowstart[1] = rowstarts[2 * i_outer + 1];
**length[1]** = rowend[1] + -1 * rowstart[1];
}
for (int j = 0; j <= -1 + **length**; ++j)
{
acc_j[0] = acc_j[0] + values[rowstart[0] + j] * x[colindices[rowstart[0] + j]];
if (-2 + -2 * i_outer + m >= 0)
acc_j[1] = acc_j[1] + values[rowstart[1] + j] * x[colindices[rowstart[1] + j]];
}
y[2 * i_outer] = acc_j[0];
if (-2 + -2 * i_outer + m >= 0)
y[2 * i_outer + 1] = acc_j[1];
}
}
```

Note the starred lines dealing with the length parameter (apparently bolding doesn't work in code markdown). The `length` parameter is correctly split by ILP _except_ for when it's used as the loop bound