I'm not sure how you got those numbers but hopefully this helps clear some things up...
The TMP36 outputs 0.75V at 25°C, with an output scale factor of 10mV/°C.
Therefore it will output 0.10V at -40°C and 1.75V at 125°C.
With a 12-bit A/D this will give an approximate reading of:
0.10V / 3.3V x 4095 = 124 decimal at -40°C
1.75V / 3.3V x 4095 = 2172 decimal at 125°C
Some problems may be arising from not having a decoupling capacitor across the supply pins of the temperature sensor. Read this note from the datasheet:
Note the 0.1 μF bypass capacitor on the input. This capacitor should be a ceramic type, have very short leads (surface-mount is preferable), and be located as close as possible in physical proximity to the temperature sensor supply pin. Because these temperature sensors operate on very little supply current and may be exposed to very hostile electrical environments, it is important to minimize the effects of radio frequency interference (RFI) on these devices. The effect of RFI on these temperature sensors specifically and on analog ICs in general is manifested as abnormal dc shifts in the output voltage due to the rectification of the high frequency ambient noise by the IC. When the devices are operated in the presence of high frequency radiated or conducted noise, a large value tantalum capacitor (±2.2 μF) placed across the 0.1 μF ceramic capacitor may offer additional noise immunity.
We are basically sticking the temp sensor right next to the Wifi chip antenna, without a decoupling capacitor...
Anyone having problems with readings should try to put a 0.1uF (or larger) ceramic cap directly between pins 1 and 3 of the TMP36.