I do not have a complete answer to your question but here are some thoughts and references. First of all, it is crucial to specify the model of computation and the representation of the input, I expect this post will clarify why. I know of essentially two cases:
- the input matrix has coefficients that are algebraic numbers (*): then the eigenvalues are also algebraic numbers and can be determined effectively, one can also determine the multiplicities of the eigenvalues. There exists very good algorithms to get rational approximations of algebraic numbers such a  where the bit complexity is $O((b+n)n^2)$ (neglecting some log factors) where $b$ is the number of bits and $n$ is the degree of the polynomial that represents the algebraic number (which in turns depends polynomially on the heigh of the coefficient of the matrix).
- the input matrix has coefficients that can be arbitrary real numbers but we can query arbitrary rational approximations of them: this fits into the theory of Computable Analysis and things get complicated. It is known that eigenvalues are uniformly computable (because they are roots of the characteristic polynomial) from the input coefficient , the algorithm runs in polynomial time (in the precision), probably using  it runs in linear time (if we ignore the cost of computing approximations of the input). However the eigenvectors are not computable nor are multiplicities, unless you known them in advance .
- your do not care about the representation of numbers and assume all arithmetic operations have a unit cost: this is the BSS model and  gives a very precise arithmetic complexity that is logarithmic in the precision (ie $O(\log b)$ where $b$ is the number of bit). But of course this completely neglects the complexity of representing and operating on the numbers and assume that equality is a unit operation (whereas it is uncomputable in general). I personally find this model to be very inadequate to express the complexity of this problem.
Based on your question, I suspect your matrix has rational inputs (which are algebraic numbers of degree 1), thus it would fit in the first case. In this case, I think (but have not checked) that the bit complexity will be linear (maybe with polylogarithmic factors) in the precision, but the constant will depend on the size and numbers in the input matrix (most probably in a polynomial fashion).
(*) algebraic numbers can be represented exactly with a finite number of bits, all arithmetic operations (and more) on algebraic numbers can be performed effectively, however the complexity details on the height of the numbers, thus on field extension in which the coefficients of the input lie.
 Optimal and nearly optimal algorithms for approximating polynomial zeros, Pan, 1996
 Complexity Theory of Real Functions, Ko
 A Computable Spectral Theorem, Brattka and Ziegler
 The Complexity of the Matrix Eigenproblem, Pan and Chen