search::par_search_blocking: Remember node ids for matching ways and
return the positions of these nodes in a second pass over the file.

* Decode/search in parallel
* Start decoding the first Blob as soon as possible
* Update multiple markers at once

Use the parameters -s and --pbf for this.

* Cache the enabled vertex attributes in `Context`
* The vertex attribute parameters (offset, stride, ...) are not cached,
  yet
* Store parameters as `VertexAttribParams` in `Program`

This reduces calls to glUseBuffer.

This would reduce the calls to glUseProgram if there actually was more
than one program.

Each Texture is bound to a unique texture unit (glBindTexture is called
only once). Context keeps track of the current active texture unit and
reduces the number of glActiveTexture calls to a minimum.

Program, Buffer, Texture... structs no longer store a reference to a
Context. Now it has to be passed explicitly to all methods that mutate
the OpenGL state. This enables caching the state and removing redundant
calls to gl* functions.

and only call glViewport once per frame when resizing.

This should improve platform independence.