For the time being there is no workaround.
HTML-to-PDF conversion process consists of 2 phases: Initially PD4ML renders HTML document layout for a “screen device”, whose height is unlimited. At this point page breaks are not taken into account. After that the rendered content is “printed” to PDF output device (or Image device, or RTF device etc) and depending on a particular output media nature and dimensions, page breaks (or “page-break-inside: avoid” directives) applied.
What you actually need is to reflow the text depending on an actual image position, and the actual image position depends on a page break position. So during the phase 1 we need a position info, which will be known only during the phase 2. That is a very big challenge to the current PD4ML architecture and we are even not ready to estimate needed efforts yet.