There is no inherent problem with using stl() to deseasonalize count data. One issue to be aware of however is that count data generally has an increasing variance as the mean increases. This is often seen in both the seasonal and random elements of the decomposition. Using stl() on the raw data will not take this into account, and hence it may be best to first take the logarithm (edit - or square root) of your data.
It doesn't matter that the trend values are not integers any more. They can be thought of in a similar way to the parameter in a Poisson distribution. Although a Poisson distributed variable must be an integer, the mean doesn't need to be.
However, this doesn't necessarily mean you can use lm() to model the trend component. There are many pitfalls in modelling trends in time series, as spurious correlations will be very difficult to avoid. More commonly people first detrend the series and then model the residual part.