For instance, Cagetti (2003) estimates ρ > 2 by targeting the median wealth of households, Gourinchas and Parker (2002) find ρ < 2 by targeting the mean consumption of households, Chetty (2006), using the effects of wage changes on labor supply, finds that ρ < 2.
Marginal utility of consumption is decreasing very fast when ρ > 2, still you can find in macro or finance even ρ = 10. In experimental or in labor ρ is close to 0.
Is there any bound on this coefficient?
A practical or mathematical bound? For a practical bound, people have argued that very large values of $\rho$ are impossible to reconcile with non-financial risk taking. Indeed, some values of $\rho$ may not be compatible with crossing the street.
Perhaps that's a bit glib, but consider Neilson and Winter (2001) which find that CRRA utility and observed wage premia for dangerous jobs are compatible with values of $\rho$ such that $0<\rho<<1$. They then point out that if $\rho$ were very very large and the model properly specified then the dangerous job wage premium would also be large or alternatively (and implausibly) that people place a very, very low value on their own lives (like $3,000).
Generally speaking, data and models implying very high values of $\rho$ like a basic equity premium setup and the relative volatility of equity returns and consumption) are often taken as signs of model miss-specification instead of alternative measures of risk preferences.
While we can't talk about globally infinite risk aversion, we can consider what happens as If you take the limit $\rho\rightarrow\infty$ around a particular value of wealth $W_0$. As it goes to infinity you get a step-function utility function with $U(W) = -\infty, W < W_0$ and $U(W) = C, W \ge W_0$.
In the benchmark model of representative household intertemporal utility maximization problem (the "Ramsey model", see for example, Barro & Sala-I-Martin (2004), Economic Growth (2n ed), ch. 2, using a Constant Relative Risk Aversion utility function
$$u(c) = \frac {c^{1-\theta}}{1-\theta}$$
results in the optimal rule for per capita consumption growth $$\frac {\dot c}{c} = \frac 1{\theta} (r-\rho)$$
where $r$ is the net rate of return on assets and $\rho$ the rate of pure time preference.
$\theta$ is the coefficient of relative risk aversion (for which the OP uses $\rho$ in the question).
Benchmark values for post-war western economies are $r=0.06$ and $\rho=0.02$. So we would have
$$\frac {\dot c}{c} = \frac {0.04}{\theta}$$
Therefore for $\theta =2$ we get $\frac {\dot c}{c}=0.02$ which is consistent with historical data, while for $\theta = 10$ we would get $\frac {\dot c}{c}=0.004$.
In another aspect in this model, the coefficient of relative risk aversion is inversely related to the gross savings rate, so a very-high value of $\theta$ would imply a counter-factually low steady-sate savings rate (where "gross savings rate" should be interpreted broadly, as any form of deferred consumption, including investments in human capital, something that also implies that the concept of "capital" and its share should be extended to include "human capital" also).
