[depthMap] Compute depth map with a single stream per tile

[gregoire-dl]

This PR is based on https://github.com/alicevision/AliceVision/pull/1155 The objective of this PR is to compute depth/similarity map per tile using one CUDA stream per tile.

Features list

• [x] Remove legacy Refine strategy.
• [x] Remove DepthSimMap class
• [x] Use a CUDA kernel to upscale a depth/similarity map.
• [x] Use a CUDA kernel to compute pixSize from a depth/similarity map.
• [x] Use fixed allocations for Sgm & Refine objects

Implementation remarks

This PR must be tested intensively before being merged. Tests must include a multi-GPU test.