I tried making one of these a few years back. I've just had a look and I can't see it, I'll post photos if I can find it.
Dual heaters and sensors are the way to go, you want excellent isolation between the two and
slightly lower isolation between each and ambient.
I used a sandwich of two aluminium plates and two sheets of expanded polystyrene with the heater/sensors squeezed between them. I used 1206 resistors (200R) glued to the back of BAS16s (SOT-23). Connections were very fine 'magnet' wire (loosely twisted) inputs from one side, outputs from the other.
Although I was only using an LM358, results were pretty good. Zero and tracking were ok, but I got horrible overshoot (hadn't worked out the opamp time constants correctly). I dropped it in favor of some other project and never got back to it.
I've been thinking of having another go (your post is well timed) and have purchased some 1k2 0.1% 1206 resistors and BAV99s. The idea is to sandwich the BAV99 between two paralleled 1k2s to make 600R in. The BAV99 gives two series diodes, so twice the temperature coefficient, and connections on the same side of the SOT23 so better isolation (snip the center pin).
Two 100R resistors would make a 50R input, not sure about frequency performance, that would depend on the terminations (as I said, twisted pair in and out).
You probably want to look at Jim Williams's App note Lin Tech AN22, on the subject, he was using it for LCD CCFL backlight inverter evaluation and was using the LT1088 thermal converter. (edit:There's also) a rather good write-up somewhere comparing it with the technology used in the the HP3400A RMS meter.
http://cds.linear.com/docs/en/application-note/an22.pdf