Performance improvement in libcudf case conversion for long strings

[davidwendt]

Description

Improves logic efficiency overall strings case conversion and reworks the specialized kernels for long strings to improve parallelization within each string.
Closes #15406

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[davidwendt]

Existing benchmark show good improvement for long strings

## [0] NVIDIA RTX A6000

| row_width | num_rows |   Ref Time |   Cmp Time |           Diff |   %Diff |
|-----------|----------|------------|------------|----------------|---------|
|    32     |   4096   | 105.801 us |  98.869 us |      -6.932 us |  -6.55% |
|    64     |   4096   | 142.955 us | 138.110 us |      -4.845 us |  -3.39% |
|    128    |   4096   | 186.503 us | 211.072 us |      24.569 us |  13.17% |
|    256    |   4096   | 288.387 us | 214.724 us |     -73.663 us | -25.54% |
|    512    |   4096   | 514.895 us | 224.731 us |    -290.164 us | -56.35% |
|   1024    |   4096   |   1.007 ms | 262.894 us |    -743.903 us | -73.89% |
|   2048    |   4096   |   1.992 ms | 348.787 us |   -1643.051 us | -82.49% |
|    32     |  32768   | 109.572 us | 102.730 us |      -6.843 us |  -6.24% |
|    64     |  32768   | 147.685 us | 143.177 us |      -4.508 us |  -3.05% |
|    128    |  32768   | 237.028 us | 237.712 us |       0.684 us |   0.29% |
|    256    |  32768   | 354.450 us | 340.402 us |     -14.048 us |  -3.96% |
|    512    |  32768   | 651.191 us | 529.697 us |    -121.494 us | -18.66% |
|   1024    |  32768   |   1.365 ms | 914.432 us |    -450.780 us | -33.02% |
|   2048    |  32768   |   2.668 ms |   1.707 ms |    -961.703 us | -36.04% |
|    32     |  262144  | 192.874 us | 174.467 us |     -18.406 us |  -9.54% |
|    64     |  262144  | 401.702 us | 409.239 us |       7.538 us |   1.88% |
|    128    |  262144  |   2.059 ms | 961.064 us |   -1097.503 us | -53.31% |
|    256    |  262144  |   6.995 ms |   1.587 ms |   -5407.260 us | -77.31% |
|    512    |  262144  |  18.476 ms |   2.916 ms |  -15560.385 us | -84.22% |
|   1024    |  262144  |  41.705 ms |   6.044 ms |  -35661.092 us | -85.51% |
|   2048    |  262144  |  98.148 ms |  12.489 ms |  -85658.725 us | -87.28% |
|    32     | 2097152  | 928.088 us | 762.243 us |    -165.845 us | -17.87% |
|    64     | 2097152  |   2.161 ms |   2.022 ms |    -138.884 us |  -6.43% |
|    128    | 2097152  |  14.398 ms |   6.582 ms |   -7816.360 us | -54.29% |
|    256    | 2097152  |  55.781 ms |  11.553 ms |  -44228.492 us | -79.29% |
|    512    | 2097152  | 150.897 ms |  22.285 ms | -128611.893 us | -85.23% |
|    32     | 16777216 |   7.042 ms |   5.710 ms |   -1331.347 us | -18.91% |
|    64     | 16777216 |  16.597 ms |  15.268 ms |   -1329.738 us |  -8.01% |

[PointKernel]

probably due to the high register usage with cub

[PointKernel]

Clean as always. LGTM

[davidwendt]

/merge