It looks like the Gyro and the Magnetometer are slightly miscalibrated. In excel, it looks like they are about 135 degrees apart. Try changing your formula to be this instead:
=DEGREES(ATAN2(U3,T3))+Dcal
Where Dcal is some calibration value. It looks like the data matches up fairly well when Dcal is about 135 degrees.
You didn't say which model of IMU you got.
For the -1200F050 model, the datasheet indicates that the gyro noise level (one σ) is 0.56°/s typical, and given that the sensitivity is 0.055°/s/count, that would give you a noise level that hovers around 10 counts.
For the -0150F050 model, the noise level (one σ) is 0.36°/s typical, and given that the sensitivity is 0.0068°/s/count, that would give you a noise level that hovers around 53 counts.
I didn't analyze your data, since you didn't provide it in copyable form, but those numbers look about right (including bias) for the more sensitive model.
ADDITIONAL
Keep in mind that the noise specification is a statistical measurement. The noise is assumed to be "additive white Gaussian noise" (AWGN), and the level is given as a one-sigma (σ) value, which is a standard way of characterizing a Gaussian distribution (the standard deviation). You can think of this as a kind of RMS value; it's equivalent to specifying the noise power when you're talking about voltage or current.
What this means is that 68.2% of the time (68.2% of the samples) will have a noise value that's less than the specified number (53 counts in this case). The rest of the time, the value can be larger — in some cases, much larger.
If you want to get a better idea of what you might consider to be a "peak" value fore the noise, you need to look at the 3σ value — 99.8% of the time, the noise in a sample will be less than 3 × 53 = 159 counts.
If you can provide your data in machine-readable format, it's very straightforward to calculate its mean and standard deviation to see whether it matches the sensor specifications. I suspect that it's well within spec.
INTEGRATION
When you integrate a value that has both a bias error term and a noise term, the result will wander away from the correct value even if the underlying "signal" is zero. The bias term (essentially a constant), when integrated, becomees a linear "ramp", which means that the error grows linearly with time. The noise term, when integrated, becomes a "random walk", and the output error associated with this grows proportionally to the square root of the integration time.
Unless you can find a way to compensate for these errors (using data from other sensors), the result will become useless within minutes or even seconds, depending on the quality of the sensor.
Best Answer
The fate of BNO055 is unknown at the moment. Quite for a long time Bosch was pushing its own alternatives without pre-programmed fusion software, like BMF055. On one hand it allowed developers to make their own changes, on the other you had to pay license fees for BSX software, which I suppose was the whole point.
Now Bosch partnered with CEVA Hillcrest Labs. CEVA programs Bosch parts with their own software and sells them as BNO085. I guess this arrangement is more economically beneficial to Bosch, so BNO055 started to disappear. Unfortunately, CEVA seems to be even hungrier than Bosch, you cannot even download the datasheet without registration with them.
You may try your luck with BNO085, of course. But in the long run you may be better using one of hundreds 6- or 9-DOF sensors and running your own fusion software separately.