When computing the width of a box with `width: auto` in an intrinsic
sizing context (min-content or max-content), we previously called
calculate_min/max_content_width() for every such box.
This was wasteful for boxes that don't establish their own formatting
context (e.g. plain divs in a BFC), since their intrinsic width is
simply the intrinsic width of their contents, which we're already
computing as part of the current layout pass.
By only computing intrinsic widths separately for boxes that actually
create a new formatting context (BFC, flex, grid, table, etc.), we
avoid creating redundant throwaway LayoutState objects and formatting
contexts.
For deeply nested structures with `width: max-content` on an ancestor,
this reduces the number of formatting contexts created from O(n) to
O(1), where n is the nesting depth.
