I may also add that I don't foresee giant steps on the sensor side. They are decently close to physical limits.
Sensor seem close to physical limits as far as the quality of the signal coming out of the photosites, but there is still significant room for improvement (particularly at Canon) in minimizing noise from later stages of the processing pipeline, so that Guillermo's ideal of "ETTL" could be pursued. When the most exposure you can get to the sensor is short of blowing out any photosites, the only important constraint on choice of ISO speed setting should be avoiding amplifying any highlights into clipping; one should not have to worry about a too low ISO exposure index choice leading to further degradation of shadows beyond the noise already present in the photosites. That is, the ISO speed setting should some day be reduced to a convenience, indicating the intended level placement for the conversion from raw. Already, with some sensors it seems that any speed setting beyond about ISO 400 might as well be done with no further amplification of the analog signal or in the raw file levels, but just in conversion from raw to JPEG (or whatever final display format is used), a flag in the raw file to indicated the intended default conversion level (including display on the rear-screen or EVF.) Even super-fancy multi-million-point light metering could ideally just record information about the maximum pixel level in the raw file, rather than committing to the accuracy of that reading through extra amplification of raw levels.
... Making [the sensors] larger helps a bit, but than you need a longer lens. Twice the sensor diameter, half the noise and eight times heavier lens. To that comes the $s.
As you say at the end, for birding and other telephoto cases, this is in fact almost entirely about lenses that gather light faster, not larger sensors. More precisely, the most fundamental physical constraint on low light handling is
entrance pupil diameter: focal length divided by aperture ratio. A 36x24mm "35mm format" sensor with a 600/4 lens is great for birds, but as far as shadow handling, probably a 24x16mm "APS-C" sensor with 400/2.8 lens or a 17.3x13mm "Four Thirds" sensor with 300/2 lens would be as good, due to the huge 150mm entrance pupil diameter in each case -- and the lenses would be about as heavy and expensive too!
My guideline is that when the user of a format would benefit from a prime faster than about f/1.4 to f/2 or a zoom faster than about f/2.8, that is when a larger sensor format starts to offer a low-light handling advantage, through the easier optical design of longer lenses of the same entrance pupil size.